fix(workaround): install vulkan from repos for amd64

This is a temporary workaround to confirm that the vulkan binary
upstream are not built with the same glibc version of ubuntu 24.04,
which makes it incompatible.

See https://github.com/mudler/LocalAI/issues/7914

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

Dockerfile

@@ -16,7 +16,7 @@ RUN apt-get update && \
 
 # The requirements-drivers target is for BUILD_TYPE specific items.  If you need to install something specific to CUDA, or specific to ROCM, it goes here.
 FROM requirements AS requirements-drivers
-
+ARG VULKAN_FROM_SOURCE=false
 ARG BUILD_TYPE
 ARG CUDA_MAJOR_VERSION=12
 ARG CUDA_MINOR_VERSION=0
@@ -41,7 +41,7 @@ RUN <<EOT bash
             git python-is-python3 bison libx11-xcb-dev liblz4-dev libzstd-dev \
             ocaml-core ninja-build pkg-config libxml2-dev wayland-protocols python3-jsonschema \
             clang-format qtbase5-dev qt6-base-dev libxcb-glx0-dev sudo xz-utils mesa-vulkan-drivers
-        if [ "amd64" = "$TARGETARCH" ]; then
+        if [ "amd64" = "$TARGETARCH" ] && [ "${VULKAN_FROM_SOURCE}" = "true" ]; then
             wget "https://sdk.lunarg.com/sdk/download/1.4.328.1/linux/vulkansdk-linux-x86_64-1.4.328.1.tar.xz" && \
             tar -xf vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
             rm vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
@@ -59,6 +59,11 @@ RUN <<EOT bash
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/include/* /usr/include/ && \
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/share/* /usr/share/ && \
             rm -rf /opt/vulkan-sdk
+        elif [ "amd64" = "${TARGETARCH}}" ]; then
+            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc && \
+            wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list http://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
+            apt-get update && \
+            apt-get install -y vulkan-sdk
         fi
         if [ "arm64" = "$TARGETARCH" ]; then
             mkdir vulkan && cd vulkan && \

backend/Dockerfile.golang

@@ -46,7 +46,7 @@ RUN <<EOT bash
             git python-is-python3 bison libx11-xcb-dev liblz4-dev libzstd-dev \
             ocaml-core ninja-build pkg-config libxml2-dev wayland-protocols python3-jsonschema \
             clang-format qtbase5-dev qt6-base-dev libxcb-glx0-dev sudo xz-utils
-        if [ "amd64" = "$TARGETARCH" ]; then
+        if [ "amd64" = "$TARGETARCH" ] && [ "${VULKAN_FROM_SOURCE}" = "true" ]; then
             wget "https://sdk.lunarg.com/sdk/download/1.4.328.1/linux/vulkansdk-linux-x86_64-1.4.328.1.tar.xz" && \
             tar -xf vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
             rm vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
@@ -64,6 +64,11 @@ RUN <<EOT bash
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/include/* /usr/include/ && \
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/share/* /usr/share/ && \
             rm -rf /opt/vulkan-sdk
+        elif [ "amd64" = "${TARGETARCH}}" ]; then
+            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc && \
+            wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list http://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
+            apt-get update && \
+            apt-get install -y vulkan-sdk
         fi
         if [ "arm64" = "$TARGETARCH" ]; then
             mkdir vulkan && cd vulkan && \

backend/Dockerfile.llama-cpp

@@ -103,7 +103,7 @@ RUN <<EOT bash
             git python-is-python3 bison libx11-xcb-dev liblz4-dev libzstd-dev \
             ocaml-core ninja-build pkg-config libxml2-dev wayland-protocols python3-jsonschema \
             clang-format qtbase5-dev qt6-base-dev libxcb-glx0-dev sudo xz-utils
-        if [ "amd64" = "$TARGETARCH" ]; then
+        if [ "amd64" = "$TARGETARCH" ] && [ "${VULKAN_FROM_SOURCE}" = "true" ]; then
             wget "https://sdk.lunarg.com/sdk/download/1.4.328.1/linux/vulkansdk-linux-x86_64-1.4.328.1.tar.xz" && \
             tar -xf vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
             rm vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
@@ -121,6 +121,11 @@ RUN <<EOT bash
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/include/* /usr/include/ && \
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/share/* /usr/share/ && \
             rm -rf /opt/vulkan-sdk
+        elif [ "amd64" = "${TARGETARCH}}" ]; then
+            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc && \
+            wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list http://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
+            apt-get update && \
+            apt-get install -y vulkan-sdk
         fi
         if [ "arm64" = "$TARGETARCH" ]; then
             mkdir vulkan && cd vulkan && \

backend/Dockerfile.python

@@ -60,7 +60,7 @@ RUN <<EOT bash
             git python-is-python3 bison libx11-xcb-dev liblz4-dev libzstd-dev \
             ocaml-core ninja-build pkg-config libxml2-dev wayland-protocols python3-jsonschema \
             clang-format qtbase5-dev qt6-base-dev libxcb-glx0-dev sudo xz-utils
-        if [ "amd64" = "$TARGETARCH" ]; then
+        if [ "amd64" = "$TARGETARCH" ] && [ "${VULKAN_FROM_SOURCE}" = "true" ]; then
             wget "https://sdk.lunarg.com/sdk/download/1.4.328.1/linux/vulkansdk-linux-x86_64-1.4.328.1.tar.xz" && \
             tar -xf vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
             rm vulkansdk-linux-x86_64-1.4.328.1.tar.xz && \
@@ -78,6 +78,11 @@ RUN <<EOT bash
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/include/* /usr/include/ && \
             cp -rfv /opt/vulkan-sdk/1.4.328.1/x86_64/share/* /usr/share/ && \
             rm -rf /opt/vulkan-sdk
+        elif [ "amd64" = "${TARGETARCH}}" ]; then
+            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc && \
+            wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list http://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
+            apt-get update && \
+            apt-get install -y vulkan-sdk
         fi
         if [ "arm64" = "$TARGETARCH" ]; then
             mkdir vulkan && cd vulkan && \

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=593da7fa49503b68f9f01700be9f508f1e528992
+LLAMA_VERSION?=b1377188784f9aea26b8abde56d4aee8c733eec7
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

backend/python/chatterbox/install.sh

@@ -17,4 +17,9 @@ if [ "x${BUILD_PROFILE}" == "xintel" ]; then
 fi
 EXTRA_PIP_INSTALL_FLAGS+=" --no-build-isolation"
 
+if [ "x${BUILD_PROFILE}" == "xl4t12" ]; then
+    USE_PIP=true
+fi
+
+
 installRequirements

backend/python/chatterbox/requirements-install.txt

@@ -0,0 +1,5 @@
+# Build dependencies needed for packages installed from source (e.g., git dependencies)
+# When using --no-build-isolation, these must be installed in the venv first
+wheel
+setuptools
+packaging

backend/python/diffusers/install.sh

@@ -16,6 +16,10 @@ if [ "x${BUILD_PROFILE}" == "xintel" ]; then
     EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
 fi
 
+if [ "x${BUILD_PROFILE}" == "xl4t12" ]; then
+    USE_PIP=true
+fi
+
 # Use python 3.12 for l4t
 if [ "x${BUILD_PROFILE}" == "xl4t13" ]; then
   PYTHON_VERSION="3.12"

backend/python/kokoro/install.sh

@@ -16,4 +16,8 @@ if [ "x${BUILD_PROFILE}" == "xintel" ]; then
     EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
 fi
 
+if [ "x${BUILD_PROFILE}" == "xl4t12" ]; then
+    USE_PIP=true
+fi
+
 installRequirements

backend/python/neutts/install.sh

@@ -26,6 +26,12 @@ fi
 
 EXTRA_PIP_INSTALL_FLAGS+=" --no-build-isolation"
 
+
+if [ "x${BUILD_PROFILE}" == "xl4t12" ]; then
+    USE_PIP=true
+fi
+
+
 git clone https://github.com/neuphonic/neutts-air neutts-air
 
 cp -rfv neutts-air/neuttsair ./

backend/python/vibevoice/install.sh

@@ -23,6 +23,10 @@ if [ "x${BUILD_PROFILE}" == "xl4t13" ]; then
   PY_STANDALONE_TAG="20251120"
 fi
 
+if [ "x${BUILD_PROFILE}" == "xl4t12" ]; then
+    USE_PIP=true
+fi
+
 installRequirements
 
 git clone https://github.com/microsoft/VibeVoice.git

core/gallery/backend_types.go

@@ -63,6 +63,25 @@ func (m *GalleryBackend) IsMeta() bool {
 	return len(m.CapabilitiesMap) > 0 && m.URI == ""
 }
 
+// IsCompatibleWith checks if the backend is compatible with the current system capability.
+// For meta backends, it checks if any of the capabilities in the map match the system capability.
+// For concrete backends, it delegates to SystemState.IsBackendCompatible.
+func (m *GalleryBackend) IsCompatibleWith(systemState *system.SystemState) bool {
+	if systemState == nil {
+		return true
+	}
+
+	// Meta backends are compatible if the system capability matches one of the keys
+	if m.IsMeta() {
+		capability := systemState.Capability(m.CapabilitiesMap)
+		_, exists := m.CapabilitiesMap[capability]
+		return exists
+	}
+
+	// For concrete backends, delegate to the system package
+	return systemState.IsBackendCompatible(m.Name, m.URI)
+}
+
 func (m *GalleryBackend) SetInstalled(installed bool) {
 	m.Installed = installed
 }

core/gallery/backends_test.go

@@ -172,6 +172,252 @@ var _ = Describe("Gallery Backends", func() {
 			Expect(nilMetaBackend.IsMeta()).To(BeFalse())
 		})
 
+		It("should check IsCompatibleWith correctly for meta backends", func() {
+			metaBackend := &GalleryBackend{
+				Metadata: Metadata{
+					Name: "meta-backend",
+				},
+				CapabilitiesMap: map[string]string{
+					"nvidia":  "nvidia-backend",
+					"amd":     "amd-backend",
+					"default": "default-backend",
+				},
+			}
+
+			// Test with nil state - should be compatible
+			Expect(metaBackend.IsCompatibleWith(nil)).To(BeTrue())
+
+			// Test with NVIDIA system - should be compatible (has nvidia key)
+			nvidiaState := &system.SystemState{GPUVendor: "nvidia", VRAM: 8 * 1024 * 1024 * 1024}
+			Expect(metaBackend.IsCompatibleWith(nvidiaState)).To(BeTrue())
+
+			// Test with default (no GPU) - should be compatible (has default key)
+			defaultState := &system.SystemState{}
+			Expect(metaBackend.IsCompatibleWith(defaultState)).To(BeTrue())
+		})
+
+		Describe("IsCompatibleWith for concrete backends", func() {
+			Context("CPU backends", func() {
+				It("should be compatible on all systems", func() {
+					cpuBackend := &GalleryBackend{
+						Metadata: Metadata{
+							Name: "cpu-llama-cpp",
+						},
+						URI: "quay.io/go-skynet/local-ai-backends:latest-cpu-llama-cpp",
+					}
+					Expect(cpuBackend.IsCompatibleWith(&system.SystemState{})).To(BeTrue())
+					Expect(cpuBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Nvidia, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					Expect(cpuBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.AMD, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+				})
+			})
+
+			Context("Darwin/Metal backends", func() {
+				When("running on darwin", func() {
+					BeforeEach(func() {
+						if runtime.GOOS != "darwin" {
+							Skip("Skipping darwin-specific tests on non-darwin system")
+						}
+					})
+
+					It("should be compatible for MLX backend", func() {
+						mlxBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "mlx",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-mlx",
+						}
+						Expect(mlxBackend.IsCompatibleWith(&system.SystemState{})).To(BeTrue())
+					})
+
+					It("should be compatible for metal-llama-cpp backend", func() {
+						metalBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "metal-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-llama-cpp",
+						}
+						Expect(metalBackend.IsCompatibleWith(&system.SystemState{})).To(BeTrue())
+					})
+				})
+
+				When("running on non-darwin", func() {
+					BeforeEach(func() {
+						if runtime.GOOS == "darwin" {
+							Skip("Skipping non-darwin-specific tests on darwin system")
+						}
+					})
+
+					It("should NOT be compatible for MLX backend", func() {
+						mlxBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "mlx",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-mlx",
+						}
+						Expect(mlxBackend.IsCompatibleWith(&system.SystemState{})).To(BeFalse())
+					})
+
+					It("should NOT be compatible for metal-llama-cpp backend", func() {
+						metalBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "metal-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-llama-cpp",
+						}
+						Expect(metalBackend.IsCompatibleWith(&system.SystemState{})).To(BeFalse())
+					})
+				})
+			})
+
+			Context("NVIDIA/CUDA backends", func() {
+				When("running on non-darwin", func() {
+					BeforeEach(func() {
+						if runtime.GOOS == "darwin" {
+							Skip("Skipping CUDA tests on darwin system")
+						}
+					})
+
+					It("should NOT be compatible without nvidia GPU", func() {
+						cudaBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "cuda12-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-12-llama-cpp",
+						}
+						Expect(cudaBackend.IsCompatibleWith(&system.SystemState{})).To(BeFalse())
+						Expect(cudaBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.AMD, VRAM: 8 * 1024 * 1024 * 1024})).To(BeFalse())
+					})
+
+					It("should be compatible with nvidia GPU", func() {
+						cudaBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "cuda12-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-12-llama-cpp",
+						}
+						Expect(cudaBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Nvidia, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					})
+
+					It("should be compatible with cuda13 backend on nvidia GPU", func() {
+						cuda13Backend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "cuda13-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-13-llama-cpp",
+						}
+						Expect(cuda13Backend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Nvidia, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					})
+				})
+			})
+
+			Context("AMD/ROCm backends", func() {
+				When("running on non-darwin", func() {
+					BeforeEach(func() {
+						if runtime.GOOS == "darwin" {
+							Skip("Skipping AMD/ROCm tests on darwin system")
+						}
+					})
+
+					It("should NOT be compatible without AMD GPU", func() {
+						rocmBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "rocm-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-rocm-hipblas-llama-cpp",
+						}
+						Expect(rocmBackend.IsCompatibleWith(&system.SystemState{})).To(BeFalse())
+						Expect(rocmBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Nvidia, VRAM: 8 * 1024 * 1024 * 1024})).To(BeFalse())
+					})
+
+					It("should be compatible with AMD GPU", func() {
+						rocmBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "rocm-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-rocm-hipblas-llama-cpp",
+						}
+						Expect(rocmBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.AMD, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					})
+
+					It("should be compatible with hipblas backend on AMD GPU", func() {
+						hipBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "hip-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-hip-llama-cpp",
+						}
+						Expect(hipBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.AMD, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					})
+				})
+			})
+
+			Context("Intel/SYCL backends", func() {
+				When("running on non-darwin", func() {
+					BeforeEach(func() {
+						if runtime.GOOS == "darwin" {
+							Skip("Skipping Intel/SYCL tests on darwin system")
+						}
+					})
+
+					It("should NOT be compatible without Intel GPU", func() {
+						intelBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "intel-sycl-f16-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-intel-sycl-f16-llama-cpp",
+						}
+						Expect(intelBackend.IsCompatibleWith(&system.SystemState{})).To(BeFalse())
+						Expect(intelBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Nvidia, VRAM: 8 * 1024 * 1024 * 1024})).To(BeFalse())
+					})
+
+					It("should be compatible with Intel GPU", func() {
+						intelBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "intel-sycl-f16-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-intel-sycl-f16-llama-cpp",
+						}
+						Expect(intelBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Intel, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					})
+
+					It("should be compatible with intel-sycl-f32 backend on Intel GPU", func() {
+						intelF32Backend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "intel-sycl-f32-llama-cpp",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-intel-sycl-f32-llama-cpp",
+						}
+						Expect(intelF32Backend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Intel, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					})
+
+					It("should be compatible with intel-transformers backend on Intel GPU", func() {
+						intelTransformersBackend := &GalleryBackend{
+							Metadata: Metadata{
+								Name: "intel-transformers",
+							},
+							URI: "quay.io/go-skynet/local-ai-backends:latest-intel-transformers",
+						}
+						Expect(intelTransformersBackend.IsCompatibleWith(&system.SystemState{GPUVendor: system.Intel, VRAM: 8 * 1024 * 1024 * 1024})).To(BeTrue())
+					})
+				})
+			})
+
+			Context("Vulkan backends", func() {
+				It("should be compatible on CPU-only systems", func() {
+					// Vulkan backends don't have a specific GPU vendor requirement in the current logic
+					// They are compatible if no other GPU-specific pattern matches
+					vulkanBackend := &GalleryBackend{
+						Metadata: Metadata{
+							Name: "vulkan-llama-cpp",
+						},
+						URI: "quay.io/go-skynet/local-ai-backends:latest-gpu-vulkan-llama-cpp",
+					}
+					// Vulkan doesn't have vendor-specific filtering in current implementation
+					Expect(vulkanBackend.IsCompatibleWith(&system.SystemState{})).To(BeTrue())
+				})
+			})
+		})
+
 		It("should find best backend from meta based on system capabilities", func() {
 
 			metaBackend := &GalleryBackend{

core/gallery/gallery.go

@@ -226,6 +226,16 @@ func AvailableGalleryModels(galleries []config.Gallery, systemState *system.Syst
 
 // List available backends
 func AvailableBackends(galleries []config.Gallery, systemState *system.SystemState) (GalleryElements[*GalleryBackend], error) {
+	return availableBackendsWithFilter(galleries, systemState, true)
+}
+
+// AvailableBackendsUnfiltered returns all available backends without filtering by system capability.
+func AvailableBackendsUnfiltered(galleries []config.Gallery, systemState *system.SystemState) (GalleryElements[*GalleryBackend], error) {
+	return availableBackendsWithFilter(galleries, systemState, false)
+}
+
+// availableBackendsWithFilter is a helper function that lists available backends with optional filtering.
+func availableBackendsWithFilter(galleries []config.Gallery, systemState *system.SystemState, filterByCapability bool) (GalleryElements[*GalleryBackend], error) {
 	var backends []*GalleryBackend
 
 	systemBackends, err := ListSystemBackends(systemState)
@@ -241,7 +251,17 @@ func AvailableBackends(galleries []config.Gallery, systemState *system.SystemSta
 		if err != nil {
 			return nil, err
 		}
-		backends = append(backends, galleryBackends...)
+
+		// Filter backends by system capability if requested
+		if filterByCapability {
+			for _, backend := range galleryBackends {
+				if backend.IsCompatibleWith(systemState) {
+					backends = append(backends, backend)
+				}
+			}
+		} else {
+			backends = append(backends, galleryBackends...)
+		}
 	}
 
 	return backends, nil

core/http/app.go

@@ -205,6 +205,7 @@ func API(application *application.Application) (*echo.Echo, error) {
 
 	routes.RegisterLocalAIRoutes(e, requestExtractor, application.ModelConfigLoader(), application.ModelLoader(), application.ApplicationConfig(), application.GalleryService(), opcache, application.TemplatesEvaluator(), application)
 	routes.RegisterOpenAIRoutes(e, requestExtractor, application)
+	routes.RegisterAnthropicRoutes(e, requestExtractor, application)
 	if !application.ApplicationConfig().DisableWebUI {
 		routes.RegisterUIAPIRoutes(e, application.ModelConfigLoader(), application.ModelLoader(), application.ApplicationConfig(), application.GalleryService(), opcache, application)
 		routes.RegisterUIRoutes(e, application.ModelConfigLoader(), application.ModelLoader(), application.ApplicationConfig(), application.GalleryService())

core/http/endpoints/anthropic/messages.go

@@ -0,0 +1,537 @@
+package anthropic
+
+import (
+	"encoding/json"
+	"fmt"
+
+	"github.com/google/uuid"
+	"github.com/labstack/echo/v4"
+	"github.com/mudler/LocalAI/core/backend"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/http/middleware"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/templates"
+	"github.com/mudler/LocalAI/pkg/functions"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/xlog"
+)
+
+// MessagesEndpoint is the Anthropic Messages API endpoint
+// https://docs.anthropic.com/claude/reference/messages_post
+// @Summary Generate a message response for the given messages and model.
+// @Param request body schema.AnthropicRequest true "query params"
+// @Success 200 {object} schema.AnthropicResponse "Response"
+// @Router /v1/messages [post]
+func MessagesEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator *templates.Evaluator, appConfig *config.ApplicationConfig) echo.HandlerFunc {
+	return func(c echo.Context) error {
+		id := uuid.New().String()
+
+		input, ok := c.Get(middleware.CONTEXT_LOCALS_KEY_LOCALAI_REQUEST).(*schema.AnthropicRequest)
+		if !ok || input.Model == "" {
+			return sendAnthropicError(c, 400, "invalid_request_error", "model is required")
+		}
+
+		cfg, ok := c.Get(middleware.CONTEXT_LOCALS_KEY_MODEL_CONFIG).(*config.ModelConfig)
+		if !ok || cfg == nil {
+			return sendAnthropicError(c, 400, "invalid_request_error", "model configuration not found")
+		}
+
+		if input.MaxTokens <= 0 {
+			return sendAnthropicError(c, 400, "invalid_request_error", "max_tokens is required and must be greater than 0")
+		}
+
+		xlog.Debug("Anthropic Messages endpoint configuration read", "config", cfg)
+
+		// Convert Anthropic messages to OpenAI format for internal processing
+		openAIMessages := convertAnthropicToOpenAIMessages(input)
+
+		// Convert Anthropic tools to internal Functions format
+		funcs, shouldUseFn := convertAnthropicTools(input, cfg)
+
+		// Create an OpenAI-compatible request for internal processing
+		openAIReq := &schema.OpenAIRequest{
+			PredictionOptions: schema.PredictionOptions{
+				BasicModelRequest: schema.BasicModelRequest{Model: input.Model},
+				Temperature:       input.Temperature,
+				TopK:              input.TopK,
+				TopP:              input.TopP,
+				Maxtokens:         &input.MaxTokens,
+			},
+			Messages: openAIMessages,
+			Stream:   input.Stream,
+			Context:  input.Context,
+			Cancel:   input.Cancel,
+		}
+
+		// Set stop sequences
+		if len(input.StopSequences) > 0 {
+			openAIReq.Stop = input.StopSequences
+		}
+
+		// Merge config settings
+		if input.Temperature != nil {
+			cfg.Temperature = input.Temperature
+		}
+		if input.TopK != nil {
+			cfg.TopK = input.TopK
+		}
+		if input.TopP != nil {
+			cfg.TopP = input.TopP
+		}
+		cfg.Maxtokens = &input.MaxTokens
+		if len(input.StopSequences) > 0 {
+			cfg.StopWords = append(cfg.StopWords, input.StopSequences...)
+		}
+
+		// Template the prompt with tools if available
+		predInput := evaluator.TemplateMessages(*openAIReq, openAIReq.Messages, cfg, funcs, shouldUseFn)
+		xlog.Debug("Anthropic Messages - Prompt (after templating)", "prompt", predInput)
+
+		if input.Stream {
+			return handleAnthropicStream(c, id, input, cfg, ml, predInput, openAIReq, funcs, shouldUseFn)
+		}
+
+		return handleAnthropicNonStream(c, id, input, cfg, ml, predInput, openAIReq, funcs, shouldUseFn)
+	}
+}
+
+func handleAnthropicNonStream(c echo.Context, id string, input *schema.AnthropicRequest, cfg *config.ModelConfig, ml *model.ModelLoader, predInput string, openAIReq *schema.OpenAIRequest, funcs functions.Functions, shouldUseFn bool) error {
+	images := []string{}
+	for _, m := range openAIReq.Messages {
+		images = append(images, m.StringImages...)
+	}
+
+	predFunc, err := backend.ModelInference(
+		input.Context, predInput, openAIReq.Messages, images, nil, nil, ml, cfg, nil, nil, nil, "", "", nil, nil, nil)
+	if err != nil {
+		xlog.Error("Anthropic model inference failed", "error", err)
+		return sendAnthropicError(c, 500, "api_error", fmt.Sprintf("model inference failed: %v", err))
+	}
+
+	prediction, err := predFunc()
+	if err != nil {
+		xlog.Error("Anthropic prediction failed", "error", err)
+		return sendAnthropicError(c, 500, "api_error", fmt.Sprintf("prediction failed: %v", err))
+	}
+
+	result := backend.Finetune(*cfg, predInput, prediction.Response)
+	
+	// Check if the result contains tool calls
+	toolCalls := functions.ParseFunctionCall(result, cfg.FunctionsConfig)
+	
+	var contentBlocks []schema.AnthropicContentBlock
+	var stopReason string
+	
+	if shouldUseFn && len(toolCalls) > 0 {
+		// Model wants to use tools
+		stopReason = "tool_use"
+		for _, tc := range toolCalls {
+			// Parse arguments as JSON
+			var inputArgs map[string]interface{}
+			if err := json.Unmarshal([]byte(tc.Arguments), &inputArgs); err != nil {
+				xlog.Warn("Failed to parse tool call arguments as JSON", "error", err, "args", tc.Arguments)
+				inputArgs = map[string]interface{}{"raw": tc.Arguments}
+			}
+			
+			contentBlocks = append(contentBlocks, schema.AnthropicContentBlock{
+				Type:  "tool_use",
+				ID:    fmt.Sprintf("toolu_%s_%d", id, len(contentBlocks)),
+				Name:  tc.Name,
+				Input: inputArgs,
+			})
+		}
+		
+		// Add any text content before the tool calls
+		textContent := functions.ParseTextContent(result, cfg.FunctionsConfig)
+		if textContent != "" {
+			// Prepend text block
+			contentBlocks = append([]schema.AnthropicContentBlock{{Type: "text", Text: textContent}}, contentBlocks...)
+		}
+	} else {
+		// Normal text response
+		stopReason = "end_turn"
+		contentBlocks = []schema.AnthropicContentBlock{
+			{Type: "text", Text: result},
+		}
+	}
+
+	resp := &schema.AnthropicResponse{
+		ID:         fmt.Sprintf("msg_%s", id),
+		Type:       "message",
+		Role:       "assistant",
+		Model:      input.Model,
+		StopReason: &stopReason,
+		Content:    contentBlocks,
+		Usage: schema.AnthropicUsage{
+			InputTokens:  prediction.Usage.Prompt,
+			OutputTokens: prediction.Usage.Completion,
+		},
+	}
+
+	if respData, err := json.Marshal(resp); err == nil {
+		xlog.Debug("Anthropic Response", "response", string(respData))
+	}
+
+	return c.JSON(200, resp)
+}
+
+func handleAnthropicStream(c echo.Context, id string, input *schema.AnthropicRequest, cfg *config.ModelConfig, ml *model.ModelLoader, predInput string, openAIReq *schema.OpenAIRequest, funcs functions.Functions, shouldUseFn bool) error {
+	c.Response().Header().Set("Content-Type", "text/event-stream")
+	c.Response().Header().Set("Cache-Control", "no-cache")
+	c.Response().Header().Set("Connection", "keep-alive")
+
+	// Create OpenAI messages for inference
+	openAIMessages := openAIReq.Messages
+
+	images := []string{}
+	for _, m := range openAIMessages {
+		images = append(images, m.StringImages...)
+	}
+
+	// Send message_start event
+	messageStart := schema.AnthropicStreamEvent{
+		Type: "message_start",
+		Message: &schema.AnthropicStreamMessage{
+			ID:      fmt.Sprintf("msg_%s", id),
+			Type:    "message",
+			Role:    "assistant",
+			Content: []schema.AnthropicContentBlock{},
+			Model:   input.Model,
+			Usage:   schema.AnthropicUsage{InputTokens: 0, OutputTokens: 0},
+		},
+	}
+	sendAnthropicSSE(c, messageStart)
+
+	// Track accumulated content for tool call detection
+	accumulatedContent := ""
+	currentBlockIndex := 0
+	inToolCall := false
+	toolCallsEmitted := 0
+	
+	// Send initial content_block_start event
+	contentBlockStart := schema.AnthropicStreamEvent{
+		Type:         "content_block_start",
+		Index:        currentBlockIndex,
+		ContentBlock: &schema.AnthropicContentBlock{Type: "text", Text: ""},
+	}
+	sendAnthropicSSE(c, contentBlockStart)
+
+	// Stream content deltas
+	tokenCallback := func(token string, usage backend.TokenUsage) bool {
+		accumulatedContent += token
+		
+		// If we're using functions, try to detect tool calls incrementally
+		if shouldUseFn {
+			cleanedResult := functions.CleanupLLMResult(accumulatedContent, cfg.FunctionsConfig)
+			
+			// Try parsing for tool calls
+			toolCalls := functions.ParseFunctionCall(cleanedResult, cfg.FunctionsConfig)
+			
+			// If we detected new tool calls and haven't emitted them yet
+			if len(toolCalls) > toolCallsEmitted {
+				// Stop the current text block if we were in one
+				if !inToolCall && currentBlockIndex == 0 {
+					sendAnthropicSSE(c, schema.AnthropicStreamEvent{
+						Type:  "content_block_stop",
+						Index: currentBlockIndex,
+					})
+					currentBlockIndex++
+					inToolCall = true
+				}
+				
+				// Emit new tool calls
+				for i := toolCallsEmitted; i < len(toolCalls); i++ {
+					tc := toolCalls[i]
+					
+					// Send content_block_start for tool_use
+					sendAnthropicSSE(c, schema.AnthropicStreamEvent{
+						Type:  "content_block_start",
+						Index: currentBlockIndex,
+						ContentBlock: &schema.AnthropicContentBlock{
+							Type: "tool_use",
+							ID:   fmt.Sprintf("toolu_%s_%d", id, i),
+							Name: tc.Name,
+						},
+					})
+					
+					// Send input_json_delta with the arguments
+					sendAnthropicSSE(c, schema.AnthropicStreamEvent{
+						Type:  "content_block_delta",
+						Index: currentBlockIndex,
+						Delta: &schema.AnthropicStreamDelta{
+							Type:        "input_json_delta",
+							PartialJSON: tc.Arguments,
+						},
+					})
+					
+					// Send content_block_stop
+					sendAnthropicSSE(c, schema.AnthropicStreamEvent{
+						Type:  "content_block_stop",
+						Index: currentBlockIndex,
+					})
+					
+					currentBlockIndex++
+				}
+				toolCallsEmitted = len(toolCalls)
+				return true
+			}
+		}
+		
+		// Send regular text delta if not in tool call mode
+		if !inToolCall {
+			delta := schema.AnthropicStreamEvent{
+				Type:  "content_block_delta",
+				Index: 0,
+				Delta: &schema.AnthropicStreamDelta{
+					Type: "text_delta",
+					Text: token,
+				},
+			}
+			sendAnthropicSSE(c, delta)
+		}
+		return true
+	}
+
+	predFunc, err := backend.ModelInference(
+		input.Context, predInput, openAIMessages, images, nil, nil, ml, cfg, nil, nil, tokenCallback, "", "", nil, nil, nil)
+	if err != nil {
+		xlog.Error("Anthropic stream model inference failed", "error", err)
+		return sendAnthropicError(c, 500, "api_error", fmt.Sprintf("model inference failed: %v", err))
+	}
+
+	prediction, err := predFunc()
+	if err != nil {
+		xlog.Error("Anthropic stream prediction failed", "error", err)
+		return sendAnthropicError(c, 500, "api_error", fmt.Sprintf("prediction failed: %v", err))
+	}
+
+	// Send content_block_stop event for last block if we didn't close it yet
+	if !inToolCall {
+		contentBlockStop := schema.AnthropicStreamEvent{
+			Type:  "content_block_stop",
+			Index: 0,
+		}
+		sendAnthropicSSE(c, contentBlockStop)
+	}
+
+	// Determine stop reason
+	stopReason := "end_turn"
+	if toolCallsEmitted > 0 {
+		stopReason = "tool_use"
+	}
+
+	// Send message_delta event with stop_reason
+	messageDelta := schema.AnthropicStreamEvent{
+		Type: "message_delta",
+		Delta: &schema.AnthropicStreamDelta{
+			StopReason: &stopReason,
+		},
+		Usage: &schema.AnthropicUsage{
+			OutputTokens: prediction.Usage.Completion,
+		},
+	}
+	sendAnthropicSSE(c, messageDelta)
+
+	// Send message_stop event
+	messageStop := schema.AnthropicStreamEvent{
+		Type: "message_stop",
+	}
+	sendAnthropicSSE(c, messageStop)
+
+	return nil
+}
+
+func sendAnthropicSSE(c echo.Context, event schema.AnthropicStreamEvent) {
+	data, err := json.Marshal(event)
+	if err != nil {
+		xlog.Error("Failed to marshal SSE event", "error", err)
+		return
+	}
+	fmt.Fprintf(c.Response().Writer, "event: %s\ndata: %s\n\n", event.Type, string(data))
+	c.Response().Flush()
+}
+
+func sendAnthropicError(c echo.Context, statusCode int, errorType, message string) error {
+	resp := schema.AnthropicErrorResponse{
+		Type: "error",
+		Error: schema.AnthropicError{
+			Type:    errorType,
+			Message: message,
+		},
+	}
+	return c.JSON(statusCode, resp)
+}
+
+func convertAnthropicToOpenAIMessages(input *schema.AnthropicRequest) []schema.Message {
+	var messages []schema.Message
+
+	// Add system message if present
+	if input.System != "" {
+		messages = append(messages, schema.Message{
+			Role:          "system",
+			StringContent: input.System,
+			Content:       input.System,
+		})
+	}
+
+	// Convert Anthropic messages to OpenAI format
+	for _, msg := range input.Messages {
+		openAIMsg := schema.Message{
+			Role: msg.Role,
+		}
+
+		// Handle content (can be string or array of content blocks)
+		switch content := msg.Content.(type) {
+		case string:
+			openAIMsg.StringContent = content
+			openAIMsg.Content = content
+		case []interface{}:
+			// Handle array of content blocks
+			var textContent string
+			var stringImages []string
+			var toolCalls []schema.ToolCall
+			toolCallIndex := 0
+
+			for _, block := range content {
+				if blockMap, ok := block.(map[string]interface{}); ok {
+					blockType, _ := blockMap["type"].(string)
+					switch blockType {
+					case "text":
+						if text, ok := blockMap["text"].(string); ok {
+							textContent += text
+						}
+					case "image":
+						// Handle image content
+						if source, ok := blockMap["source"].(map[string]interface{}); ok {
+							if sourceType, ok := source["type"].(string); ok && sourceType == "base64" {
+								if data, ok := source["data"].(string); ok {
+									mediaType, _ := source["media_type"].(string)
+									// Format as data URI
+									dataURI := fmt.Sprintf("data:%s;base64,%s", mediaType, data)
+									stringImages = append(stringImages, dataURI)
+								}
+							}
+						}
+					case "tool_use":
+						// Convert tool_use to ToolCall format
+						toolID, _ := blockMap["id"].(string)
+						toolName, _ := blockMap["name"].(string)
+						toolInput := blockMap["input"]
+						
+						// Serialize input to JSON string
+						inputJSON, err := json.Marshal(toolInput)
+						if err != nil {
+							xlog.Warn("Failed to marshal tool input", "error", err)
+							inputJSON = []byte("{}")
+						}
+						
+						toolCalls = append(toolCalls, schema.ToolCall{
+							Index: toolCallIndex,
+							ID:    toolID,
+							Type:  "function",
+							FunctionCall: schema.FunctionCall{
+								Name:      toolName,
+								Arguments: string(inputJSON),
+							},
+						})
+						toolCallIndex++
+					case "tool_result":
+						// Convert tool_result to a message with role "tool"
+						// This is handled by creating a separate message after this block
+						// For now, we'll add it as text content
+						toolUseID, _ := blockMap["tool_use_id"].(string)
+						isError := false
+						if isErrorPtr, ok := blockMap["is_error"].(*bool); ok && isErrorPtr != nil {
+							isError = *isErrorPtr
+						}
+						
+						var resultText string
+						if resultContent, ok := blockMap["content"]; ok {
+							switch rc := resultContent.(type) {
+							case string:
+								resultText = rc
+							case []interface{}:
+								// Array of content blocks
+								for _, cb := range rc {
+									if cbMap, ok := cb.(map[string]interface{}); ok {
+										if cbMap["type"] == "text" {
+											if text, ok := cbMap["text"].(string); ok {
+												resultText += text
+											}
+										}
+									}
+								}
+							}
+						}
+						
+						// Add tool result as a tool role message
+						// We need to handle this differently - create a new message
+						if msg.Role == "user" {
+							// Store tool result info for creating separate message
+							prefix := ""
+							if isError {
+								prefix = "Error: "
+							}
+							textContent += fmt.Sprintf("\n[Tool Result for %s]: %s%s", toolUseID, prefix, resultText)
+						}
+					}
+				}
+			}
+			openAIMsg.StringContent = textContent
+			openAIMsg.Content = textContent
+			openAIMsg.StringImages = stringImages
+			
+			// Add tool calls if present
+			if len(toolCalls) > 0 {
+				openAIMsg.ToolCalls = toolCalls
+			}
+		}
+
+		messages = append(messages, openAIMsg)
+	}
+
+	return messages
+}
+
+// convertAnthropicTools converts Anthropic tools to internal Functions format
+func convertAnthropicTools(input *schema.AnthropicRequest, cfg *config.ModelConfig) (functions.Functions, bool) {
+	if len(input.Tools) == 0 {
+		return nil, false
+	}
+	
+	var funcs functions.Functions
+	for _, tool := range input.Tools {
+		f := functions.Function{
+			Name:        tool.Name,
+			Description: tool.Description,
+			Parameters:  tool.InputSchema,
+		}
+		funcs = append(funcs, f)
+	}
+	
+	// Handle tool_choice
+	if input.ToolChoice != nil {
+		switch tc := input.ToolChoice.(type) {
+		case string:
+			// "auto", "any", or "none"
+			if tc == "any" {
+				// Force the model to use one of the tools
+				cfg.SetFunctionCallString("required")
+			} else if tc == "none" {
+				// Don't use tools
+				return nil, false
+			}
+			// "auto" is the default - let model decide
+		case map[string]interface{}:
+			// Specific tool selection: {"type": "tool", "name": "tool_name"}
+			if tcType, ok := tc["type"].(string); ok && tcType == "tool" {
+				if name, ok := tc["name"].(string); ok {
+					// Force specific tool
+					cfg.SetFunctionCallString(name)
+				}
+			}
+		}
+	}
+	
+	return funcs, len(funcs) > 0 && cfg.ShouldUseFunctions()
+}

core/http/endpoints/openai/chat.go

@@ -3,6 +3,7 @@ package openai
 import (
 	"encoding/json"
 	"fmt"
+	"strings"
 	"time"
 
 	"github.com/google/uuid"
@@ -34,11 +35,54 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 			Created: created,
 			Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
 			Choices: []schema.Choice{{Delta: &schema.Message{Role: "assistant"}, Index: 0, FinishReason: nil}},
-			Object:  "chat.completion.chunk",
 		}
 		responses <- initialMessage
 
+		// Track accumulated content for reasoning extraction
+		accumulatedContent := ""
+		lastEmittedReasoning := ""
+		lastEmittedCleanedContent := ""
+
 		_, _, err := ComputeChoices(req, s, config, cl, startupOptions, loader, func(s string, c *[]schema.Choice) {}, func(s string, tokenUsage backend.TokenUsage) bool {
+			accumulatedContent += s
+			// Extract reasoning from accumulated content
+			currentReasoning, cleanedContent := functions.ExtractReasoning(accumulatedContent)
+
+			// Calculate new reasoning delta (what we haven't emitted yet)
+			var reasoningDelta *string
+			if currentReasoning != lastEmittedReasoning {
+				// Extract only the new part
+				if len(currentReasoning) > len(lastEmittedReasoning) && strings.HasPrefix(currentReasoning, lastEmittedReasoning) {
+					newReasoning := currentReasoning[len(lastEmittedReasoning):]
+					reasoningDelta = &newReasoning
+					lastEmittedReasoning = currentReasoning
+				} else if currentReasoning != "" {
+					// If reasoning changed in a non-append way, emit the full current reasoning
+					reasoningDelta = &currentReasoning
+					lastEmittedReasoning = currentReasoning
+				}
+			}
+
+			// Calculate content delta from cleaned content
+			var deltaContent string
+			if len(cleanedContent) > len(lastEmittedCleanedContent) && strings.HasPrefix(cleanedContent, lastEmittedCleanedContent) {
+				deltaContent = cleanedContent[len(lastEmittedCleanedContent):]
+				lastEmittedCleanedContent = cleanedContent
+			} else if cleanedContent != lastEmittedCleanedContent {
+				// If cleaned content changed but not in a simple append, extract delta from cleaned content
+				// This handles cases where thinking tags are removed mid-stream
+				if lastEmittedCleanedContent == "" {
+					deltaContent = cleanedContent
+					lastEmittedCleanedContent = cleanedContent
+				} else {
+					// Content changed in non-append way, use the new cleaned content
+					deltaContent = cleanedContent
+					lastEmittedCleanedContent = cleanedContent
+				}
+			}
+			// Only emit content if there's actual content (not just thinking tags)
+			// If deltaContent is empty, we still emit the response but with empty content
+
 			usage := schema.OpenAIUsage{
 				PromptTokens:     tokenUsage.Prompt,
 				CompletionTokens: tokenUsage.Completion,
@@ -49,11 +93,20 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 				usage.TimingPromptProcessing = tokenUsage.TimingPromptProcessing
 			}
 
+			delta := &schema.Message{}
+			// Only include content if there's actual content (not just thinking tags)
+			if deltaContent != "" {
+				delta.Content = &deltaContent
+			}
+			if reasoningDelta != nil && *reasoningDelta != "" {
+				delta.Reasoning = reasoningDelta
+			}
+
 			resp := schema.OpenAIResponse{
 				ID:      id,
 				Created: created,
 				Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
-				Choices: []schema.Choice{{Delta: &schema.Message{Content: &s}, Index: 0, FinishReason: nil}},
+				Choices: []schema.Choice{{Delta: delta, Index: 0, FinishReason: nil}},
 				Object:  "chat.completion.chunk",
 				Usage:   usage,
 			}
@@ -176,6 +229,10 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 		if err != nil {
 			return err
 		}
+		// Extract reasoning before processing tool calls
+		reasoning, cleanedResult := functions.ExtractReasoning(result)
+		result = cleanedResult
+
 		textContentToReturn = functions.ParseTextContent(result, config.FunctionsConfig)
 		result = functions.CleanupLLMResult(result, config.FunctionsConfig)
 		functionResults := functions.ParseFunctionCall(result, config.FunctionsConfig)
@@ -208,11 +265,20 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 				usage.TimingPromptProcessing = tokenUsage.TimingPromptProcessing
 			}
 
+			var deltaReasoning *string
+			if reasoning != "" {
+				deltaReasoning = &reasoning
+			}
+			delta := &schema.Message{Content: &result}
+			if deltaReasoning != nil {
+				delta.Reasoning = deltaReasoning
+			}
+
 			resp := schema.OpenAIResponse{
 				ID:      id,
 				Created: created,
 				Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
-				Choices: []schema.Choice{{Delta: &schema.Message{Content: &result}, Index: 0, FinishReason: nil}},
+				Choices: []schema.Choice{{Delta: delta, Index: 0, FinishReason: nil}},
 				Object:  "chat.completion.chunk",
 				Usage:   usage,
 			}
@@ -553,10 +619,18 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 		default:
 
 			tokenCallback := func(s string, c *[]schema.Choice) {
+				// Extract reasoning from the response
+				reasoning, cleanedS := functions.ExtractReasoning(s)
+				s = cleanedS
+
 				if !shouldUseFn {
 					// no function is called, just reply and use stop as finish reason
 					stopReason := FinishReasonStop
-					*c = append(*c, schema.Choice{FinishReason: &stopReason, Index: 0, Message: &schema.Message{Role: "assistant", Content: &s}})
+					message := &schema.Message{Role: "assistant", Content: &s}
+					if reasoning != "" {
+						message.Reasoning = &reasoning
+					}
+					*c = append(*c, schema.Choice{FinishReason: &stopReason, Index: 0, Message: message})
 					return
 				}
 
@@ -575,9 +649,13 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 					}
 
 					stopReason := FinishReasonStop
+					message := &schema.Message{Role: "assistant", Content: &result}
+					if reasoning != "" {
+						message.Reasoning = &reasoning
+					}
 					*c = append(*c, schema.Choice{
 						FinishReason: &stopReason,
-						Message:      &schema.Message{Role: "assistant", Content: &result}})
+						Message:      message})
 				default:
 					toolCallsReason := FinishReasonToolCalls
 					toolChoice := schema.Choice{
@@ -586,6 +664,9 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 							Role: "assistant",
 						},
 					}
+					if reasoning != "" {
+						toolChoice.Message.Reasoning = &reasoning
+					}
 
 					for _, ss := range results {
 						name, args := ss.Name, ss.Arguments
@@ -606,16 +687,20 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 						} else {
 							// otherwise we return more choices directly (deprecated)
 							functionCallReason := FinishReasonFunctionCall
+							message := &schema.Message{
+								Role:    "assistant",
+								Content: &textContentToReturn,
+								FunctionCall: map[string]interface{}{
+									"name":      name,
+									"arguments": args,
+								},
+							}
+							if reasoning != "" {
+								message.Reasoning = &reasoning
+							}
 							*c = append(*c, schema.Choice{
 								FinishReason: &functionCallReason,
-								Message: &schema.Message{
-									Role:    "assistant",
-									Content: &textContentToReturn,
-									FunctionCall: map[string]interface{}{
-										"name":      name,
-										"arguments": args,
-									},
-								},
+								Message:      message,
 							})
 						}
 					}

core/http/routes/anthropic.go

@@ -0,0 +1,108 @@
+package routes
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+
+	"github.com/google/uuid"
+	"github.com/labstack/echo/v4"
+	"github.com/mudler/LocalAI/core/application"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/http/endpoints/anthropic"
+	"github.com/mudler/LocalAI/core/http/middleware"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/xlog"
+)
+
+func RegisterAnthropicRoutes(app *echo.Echo,
+	re *middleware.RequestExtractor,
+	application *application.Application) {
+
+	// Anthropic Messages API endpoint
+	messagesHandler := anthropic.MessagesEndpoint(
+		application.ModelConfigLoader(),
+		application.ModelLoader(),
+		application.TemplatesEvaluator(),
+		application.ApplicationConfig(),
+	)
+
+	messagesMiddleware := []echo.MiddlewareFunc{
+		middleware.TraceMiddleware(application),
+		re.BuildFilteredFirstAvailableDefaultModel(config.BuildUsecaseFilterFn(config.FLAG_CHAT)),
+		re.SetModelAndConfig(func() schema.LocalAIRequest { return new(schema.AnthropicRequest) }),
+		setAnthropicRequestContext(application.ApplicationConfig()),
+	}
+
+	// Main Anthropic endpoint
+	app.POST("/v1/messages", messagesHandler, messagesMiddleware...)
+
+	// Also support without version prefix for compatibility
+	app.POST("/messages", messagesHandler, messagesMiddleware...)
+}
+
+// setAnthropicRequestContext sets up the context and cancel function for Anthropic requests
+func setAnthropicRequestContext(appConfig *config.ApplicationConfig) echo.MiddlewareFunc {
+	return func(next echo.HandlerFunc) echo.HandlerFunc {
+		return func(c echo.Context) error {
+			input, ok := c.Get(middleware.CONTEXT_LOCALS_KEY_LOCALAI_REQUEST).(*schema.AnthropicRequest)
+			if !ok || input.Model == "" {
+				return echo.NewHTTPError(http.StatusBadRequest, "model is required")
+			}
+
+			cfg, ok := c.Get(middleware.CONTEXT_LOCALS_KEY_MODEL_CONFIG).(*config.ModelConfig)
+			if !ok || cfg == nil {
+				return echo.NewHTTPError(http.StatusBadRequest, "model configuration not found")
+			}
+
+			// Extract or generate the correlation ID
+			// Anthropic uses x-request-id header
+			correlationID := c.Request().Header.Get("x-request-id")
+			if correlationID == "" {
+				correlationID = uuid.New().String()
+			}
+			c.Response().Header().Set("x-request-id", correlationID)
+
+			// Set up context with cancellation
+			reqCtx := c.Request().Context()
+			c1, cancel := context.WithCancel(appConfig.Context)
+
+			// Cancel when request context is cancelled (client disconnects)
+			go func() {
+				select {
+				case <-reqCtx.Done():
+					cancel()
+				case <-c1.Done():
+					// Already cancelled
+				}
+			}()
+
+			// Add the correlation ID to the new context
+			ctxWithCorrelationID := context.WithValue(c1, middleware.CorrelationIDKey, correlationID)
+
+			input.Context = ctxWithCorrelationID
+			input.Cancel = cancel
+
+			if cfg.Model == "" {
+				xlog.Debug("replacing empty cfg.Model with input value", "input.Model", input.Model)
+				cfg.Model = input.Model
+			}
+
+			c.Set(middleware.CONTEXT_LOCALS_KEY_LOCALAI_REQUEST, input)
+			c.Set(middleware.CONTEXT_LOCALS_KEY_MODEL_CONFIG, cfg)
+
+			// Log the Anthropic API version if provided
+			anthropicVersion := c.Request().Header.Get("anthropic-version")
+			if anthropicVersion != "" {
+				xlog.Debug("Anthropic API version", "version", anthropicVersion)
+			}
+
+			// Validate max_tokens is provided
+			if input.MaxTokens <= 0 {
+				return echo.NewHTTPError(http.StatusBadRequest, fmt.Sprintf("max_tokens is required and must be greater than 0"))
+			}
+
+			return next(c)
+		}
+	}
+}

core/http/routes/ui_api.go

@@ -617,6 +617,12 @@ func RegisterUIAPIRoutes(app *echo.Echo, cl *config.ModelConfigLoader, ml *model
 			installedBackendsCount = len(installedBackends)
 		}
 
+		// Get the detected system capability
+		detectedCapability := ""
+		if appConfig.SystemState != nil {
+			detectedCapability = appConfig.SystemState.DetectedCapability()
+		}
+
 		return c.JSON(200, map[string]interface{}{
 			"backends":           backendsJSON,
 			"repositories":       appConfig.BackendGalleries,
@@ -629,6 +635,7 @@ func RegisterUIAPIRoutes(app *echo.Echo, cl *config.ModelConfigLoader, ml *model
 			"totalPages":         totalPages,
 			"prevPage":           prevPage,
 			"nextPage":           nextPage,
+			"systemCapability":   detectedCapability,
 		})
 	})
 

core/http/static/chat.js

@@ -1368,6 +1368,7 @@ async function promptGPT(systemPrompt, input) {
     let lastAssistantMessageIndex = -1;
     let lastThinkingMessageIndex = -1;
     let lastThinkingScrollTime = 0;
+    let hasReasoningFromAPI = false; // Track if we're receiving reasoning from API (skip tag-based detection)
     const THINKING_SCROLL_THROTTLE = 200; // Throttle scrolling to every 200ms
 
     try {
@@ -1401,19 +1402,24 @@ async function promptGPT(systemPrompt, input) {
               // Handle different event types
               switch (eventData.type) {
                 case "reasoning":
+                  hasReasoningFromAPI = true; // Mark that we're receiving reasoning from API
                   if (eventData.content) {
-                    // Insert reasoning before assistant message if it exists
+                    const currentChat = chatStore.getChat(chatId);
+                    if (!currentChat) break; // Chat was deleted
+                    const isMCPMode = currentChat.mcpMode || false;
+                    const shouldExpand = !isMCPMode; // Expanded in non-MCP mode, collapsed in MCP mode
+                    // Insert thinking before assistant message if it exists (always use "thinking" role)
                     if (lastAssistantMessageIndex >= 0 && targetHistory[lastAssistantMessageIndex]?.role === "assistant") {
                       targetHistory.splice(lastAssistantMessageIndex, 0, {
-                        role: "reasoning",
+                        role: "thinking",
                         content: eventData.content,
                         html: DOMPurify.sanitize(marked.parse(eventData.content)),
                         image: [],
                         audio: [],
-                        expanded: false // Reasoning is always collapsed
+                        expanded: shouldExpand
                       });
                       lastAssistantMessageIndex++; // Adjust index since we inserted
-                      // Scroll smoothly after adding reasoning
+                      // Scroll smoothly after adding thinking
                       setTimeout(() => {
                         const chatContainer = document.getElementById('chat');
                         if (chatContainer) {
@@ -1425,7 +1431,7 @@ async function promptGPT(systemPrompt, input) {
                       }, 100);
                     } else {
                       // No assistant message yet, just add normally
-                      chatStore.add("reasoning", eventData.content, null, null, chatId);
+                      chatStore.add("thinking", eventData.content, null, null, chatId);
                     }
                   }
                   break;
@@ -1491,14 +1497,17 @@ async function promptGPT(systemPrompt, input) {
                     // Only update display if this is the active chat (interval will handle it)
                     // Don't call updateTokensPerSecond here to avoid unnecessary updates
                     
-                    // Check for thinking tags in the chunk (incremental detection)
-                    if (contentChunk.includes("<thinking>") || contentChunk.includes("<think>")) {
-                      isThinking = true;
-                      thinkingContent = "";
-                      lastThinkingMessageIndex = -1;
-                    }
-                    
-                    if (contentChunk.includes("</thinking>") || contentChunk.includes("</think>")) {
+                    // Only check for thinking tags if we're NOT receiving reasoning from API
+                    // This prevents duplicate thinking/reasoning messages
+                    if (!hasReasoningFromAPI) {
+                      // Check for thinking tags in the chunk (incremental detection)
+                      if (contentChunk.includes("<thinking>") || contentChunk.includes("<think>")) {
+                        isThinking = true;
+                        thinkingContent = "";
+                        lastThinkingMessageIndex = -1;
+                      }
+                      
+                      if (contentChunk.includes("</thinking>") || contentChunk.includes("</think>")) {
                       isThinking = false;
                       // When closing tag is detected, process the accumulated thinking content
                       if (thinkingContent.trim()) {
@@ -1552,10 +1561,11 @@ async function promptGPT(systemPrompt, input) {
                         }
                         thinkingContent = "";
                       }
+                      }
                     }
                     
-                    // Handle content based on thinking state
-                    if (isThinking) {
+                    // Handle content based on thinking state (only if not receiving reasoning from API)
+                    if (!hasReasoningFromAPI && isThinking) {
                       thinkingContent += contentChunk;
                       const currentChat = chatStore.getChat(chatId);
                       if (!currentChat) break; // Chat was deleted
@@ -1637,7 +1647,10 @@ async function promptGPT(systemPrompt, input) {
           
           // Process any thinking tags that might be in the accumulated content
           // This handles cases where tags are split across chunks
-          const { regularContent: processedRegular, thinkingContent: processedThinking } = processThinkingTags(regularContent);
+          // Only process if we're NOT receiving reasoning from API (to avoid duplicates)
+          const { regularContent: processedRegular, thinkingContent: processedThinking } = hasReasoningFromAPI
+            ? { regularContent: regularContent, thinkingContent: "" }
+            : processThinkingTags(regularContent);
           
           // Update or create assistant message with processed regular content
           const currentChat = chatStore.getChat(chatId);
@@ -1645,10 +1658,10 @@ async function promptGPT(systemPrompt, input) {
           const request = activeRequests.get(chatId);
           const requestModel = request?.model || null;
           if (lastAssistantMessageIndex === -1) {
-            if (processedRegular && processedRegular.trim()) {
-              chatStore.add("assistant", processedRegular, null, null, chatId, requestModel);
-              lastAssistantMessageIndex = targetHistory.length - 1;
-            }
+            // Create assistant message if we have any content (even if empty string after processing)
+            // This ensures the message is created and can be updated with more content later
+            chatStore.add("assistant", processedRegular || "", null, null, chatId, requestModel);
+            lastAssistantMessageIndex = targetHistory.length - 1;
           } else {
             const lastMessage = targetHistory[lastAssistantMessageIndex];
             if (lastMessage && lastMessage.role === "assistant") {
@@ -1686,7 +1699,10 @@ async function promptGPT(systemPrompt, input) {
       if (assistantContentBuffer.length > 0) {
         const regularContent = assistantContentBuffer.join("");
         // Process any remaining thinking tags that might be in the buffer
-        const { regularContent: processedRegular, thinkingContent: processedThinking } = processThinkingTags(regularContent);
+        // Only process if we're NOT receiving reasoning from API (to avoid duplicates)
+        const { regularContent: processedRegular, thinkingContent: processedThinking } = hasReasoningFromAPI
+          ? { regularContent: regularContent, thinkingContent: "" }
+          : processThinkingTags(regularContent);
         
         const currentChat = chatStore.getChat(chatId);
         if (!currentChat) {
@@ -1719,23 +1735,26 @@ async function promptGPT(systemPrompt, input) {
         }
         
         // Then update or create assistant message
+        // Always create/update assistant message if we have any content
         if (lastAssistantMessageIndex !== -1) {
           const lastMessage = targetHistory[lastAssistantMessageIndex];
           if (lastMessage && lastMessage.role === "assistant") {
             lastMessage.content = (lastMessage.content || "") + (processedRegular || "");
             lastMessage.html = DOMPurify.sanitize(marked.parse(lastMessage.content));
           }
-        } else if (processedRegular && processedRegular.trim()) {
+        } else {
+          // Create assistant message (even if empty, so it can be updated with more content)
           const request = activeRequests.get(chatId);
           const requestModel = request?.model || null;
-          chatStore.add("assistant", processedRegular, null, null, chatId, requestModel);
+          chatStore.add("assistant", processedRegular || "", null, null, chatId, requestModel);
           lastAssistantMessageIndex = targetHistory.length - 1;
         }
       }
       
       // Final thinking content flush if any data remains (from incremental detection)
+      // Only process if we're NOT receiving reasoning from API (to avoid duplicates)
       const finalChat = chatStore.getChat(chatId);
-      if (finalChat && thinkingContent.trim() && lastThinkingMessageIndex === -1) {
+      if (finalChat && !hasReasoningFromAPI && thinkingContent.trim() && lastThinkingMessageIndex === -1) {
         const finalHistory = finalChat.history;
         // Extract thinking content if tags are present
         const thinkingMatch = thinkingContent.match(/<(?:thinking|redacted_reasoning)>(.*?)<\/(?:thinking|redacted_reasoning)>/s);
@@ -1891,9 +1910,13 @@ async function promptGPT(systemPrompt, input) {
     let buffer = "";
     let contentBuffer = [];
     let thinkingContent = "";
+    let reasoningContent = ""; // Track reasoning from API reasoning field
     let isThinking = false;
     let lastThinkingMessageIndex = -1;
+    let lastReasoningMessageIndex = -1; // Track reasoning message separately
+    let lastAssistantMessageIndex = -1; // Track assistant message for reasoning placement
     let lastThinkingScrollTime = 0;
+    let hasReasoningFromAPI = false; // Track if we're receiving reasoning from API (skip tag-based detection)
     const THINKING_SCROLL_THROTTLE = 200; // Throttle scrolling to every 200ms
 
     try {
@@ -1929,30 +1952,100 @@ async function promptGPT(systemPrompt, input) {
                 chatStore.updateTokenUsage(jsonData.usage, chatId);
               }
               
-              const token = jsonData.choices[0].delta.content;
+              const token = jsonData.choices?.[0]?.delta?.content;
+              const reasoningDelta = jsonData.choices?.[0]?.delta?.reasoning;
 
-              if (token) {
-                // Check for thinking tags
-                if (token.includes("<thinking>") || token.includes("<think>")) {
-                  isThinking = true;
-                  thinkingContent = "";
-                  lastThinkingMessageIndex = -1;
+              // Handle reasoning from API reasoning field - always use "thinking" role
+              if (reasoningDelta && reasoningDelta.trim() !== "") {
+                hasReasoningFromAPI = true; // Mark that we're receiving reasoning from API
+                reasoningContent += reasoningDelta;
+                const currentChat = chatStore.getChat(chatId);
+                if (!currentChat) {
+                  // Chat was deleted, skip this line
                   return;
                 }
-                if (token.includes("</thinking>") || token.includes("</think>")) {
-                  isThinking = false;
-                  if (thinkingContent.trim()) {
-                    // Only add the final thinking message if we don't already have one
-                    if (lastThinkingMessageIndex === -1) {
-                      chatStore.add("thinking", thinkingContent, null, null, chatId);
+                const isMCPMode = currentChat.mcpMode || false;
+                const shouldExpand = !isMCPMode; // Expanded in non-MCP mode, collapsed in MCP mode
+                
+                // Only create/update thinking message if we have actual content
+                if (reasoningContent.trim() !== "") {
+                  // Update or create thinking message (always use "thinking" role, not "reasoning")
+                  if (lastReasoningMessageIndex === -1) {
+                    // Find the last assistant message index to insert thinking before it
+                    const targetHistory = currentChat.history;
+                    const assistantIndex = targetHistory.length - 1;
+                    if (assistantIndex >= 0 && targetHistory[assistantIndex]?.role === "assistant") {
+                      // Insert thinking before assistant message
+                      targetHistory.splice(assistantIndex, 0, {
+                        role: "thinking",
+                        content: reasoningContent,
+                        html: DOMPurify.sanitize(marked.parse(reasoningContent)),
+                        image: [],
+                        audio: [],
+                        expanded: shouldExpand
+                      });
+                      lastReasoningMessageIndex = assistantIndex;
+                      lastAssistantMessageIndex = assistantIndex + 1; // Adjust for inserted thinking
+                    } else {
+                      // No assistant message yet, just add normally
+                      chatStore.add("thinking", reasoningContent, null, null, chatId);
+                      lastReasoningMessageIndex = currentChat.history.length - 1;
+                    }
+                  } else {
+                    // Update existing thinking message
+                    const targetHistory = currentChat.history;
+                    if (lastReasoningMessageIndex >= 0 && lastReasoningMessageIndex < targetHistory.length) {
+                      const thinkingMessage = targetHistory[lastReasoningMessageIndex];
+                      if (thinkingMessage && thinkingMessage.role === "thinking") {
+                        thinkingMessage.content = reasoningContent;
+                        thinkingMessage.html = DOMPurify.sanitize(marked.parse(reasoningContent));
+                      }
                     }
                   }
-                  return;
                 }
+                
+                // Scroll when reasoning is updated (throttled)
+                const now = Date.now();
+                if (now - lastThinkingScrollTime > THINKING_SCROLL_THROTTLE) {
+                  lastThinkingScrollTime = now;
+                  setTimeout(() => {
+                    const chatContainer = document.getElementById('chat');
+                    if (chatContainer) {
+                      chatContainer.scrollTo({
+                        top: chatContainer.scrollHeight,
+                        behavior: 'smooth'
+                      });
+                    }
+                    scrollThinkingBoxToBottom();
+                  }, 100);
+                }
+              }
 
-                // Handle content based on thinking state
-                if (isThinking) {
-                  thinkingContent += token;
+              if (token && token.trim() !== "") {
+                // Only check for thinking tags if we're NOT receiving reasoning from API
+                // This prevents duplicate thinking/reasoning messages
+                if (!hasReasoningFromAPI) {
+                  // Check for thinking tags (legacy support - models that output tags directly)
+                  if (token.includes("<thinking>") || token.includes("<think>")) {
+                    isThinking = true;
+                    thinkingContent = "";
+                    lastThinkingMessageIndex = -1;
+                    return;
+                  }
+                  if (token.includes("</thinking>") || token.includes("</think>")) {
+                    isThinking = false;
+                    if (thinkingContent.trim()) {
+                      // Only add the final thinking message if we don't already have one
+                      if (lastThinkingMessageIndex === -1) {
+                        chatStore.add("thinking", thinkingContent, null, null, chatId);
+                      }
+                    }
+                    return;
+                  }
+
+                  // Handle content based on thinking state
+                  if (isThinking) {
+                    thinkingContent += token;
                   // Count tokens for rate calculation (per chat)
                   const request = activeRequests.get(chatId);
                   if (request) {
@@ -1995,7 +2088,42 @@ async function promptGPT(systemPrompt, input) {
                     }, 100);
                   }
                 } else {
+                  // Not in thinking state, add to content buffer
                   contentBuffer.push(token);
+                  // Track assistant message index for reasoning placement
+                  if (lastAssistantMessageIndex === -1) {
+                    const currentChat = chatStore.getChat(chatId);
+                    if (currentChat) {
+                      const targetHistory = currentChat.history;
+                      // Find or create assistant message index
+                      for (let i = targetHistory.length - 1; i >= 0; i--) {
+                        if (targetHistory[i].role === "assistant") {
+                          lastAssistantMessageIndex = i;
+                          break;
+                        }
+                      }
+                      // If no assistant message yet, it will be created when we flush contentBuffer
+                    }
+                  }
+                }
+                } else {
+                  // Receiving reasoning from API, just add token to content buffer
+                  contentBuffer.push(token);
+                  // Track assistant message index for reasoning placement
+                  if (lastAssistantMessageIndex === -1) {
+                    const currentChat = chatStore.getChat(chatId);
+                    if (currentChat) {
+                      const targetHistory = currentChat.history;
+                      // Find or create assistant message index
+                      for (let i = targetHistory.length - 1; i >= 0; i--) {
+                        if (targetHistory[i].role === "assistant") {
+                          lastAssistantMessageIndex = i;
+                          break;
+                        }
+                      }
+                      // If no assistant message yet, it will be created when we flush contentBuffer
+                    }
+                  }
                 }
               }
             } catch (error) {
@@ -2007,6 +2135,17 @@ async function promptGPT(systemPrompt, input) {
         // Efficiently update the chat in batch
         if (contentBuffer.length > 0) {
           addToChat(contentBuffer.join(""));
+          // Update assistant message index after adding content
+          const currentChat = chatStore.getChat(chatId);
+          if (currentChat) {
+            const targetHistory = currentChat.history;
+            for (let i = targetHistory.length - 1; i >= 0; i--) {
+              if (targetHistory[i].role === "assistant") {
+                lastAssistantMessageIndex = i;
+                break;
+              }
+            }
+          }
           contentBuffer = [];
           // Scroll when assistant content is updated (this will also show thinking messages above)
           setTimeout(() => {
@@ -2025,7 +2164,30 @@ async function promptGPT(systemPrompt, input) {
       if (contentBuffer.length > 0) {
         addToChat(contentBuffer.join(""));
       }
+      
+      // Final reasoning flush if any data remains - always use "thinking" role
       const finalChat = chatStore.getChat(chatId);
+      if (finalChat && reasoningContent.trim() && lastReasoningMessageIndex === -1) {
+        const isMCPMode = finalChat.mcpMode || false;
+        const shouldExpand = !isMCPMode;
+        const targetHistory = finalChat.history;
+        // Find assistant message to insert before
+        const assistantIndex = targetHistory.length - 1;
+        if (assistantIndex >= 0 && targetHistory[assistantIndex]?.role === "assistant") {
+          targetHistory.splice(assistantIndex, 0, {
+            role: "thinking",
+            content: reasoningContent,
+            html: DOMPurify.sanitize(marked.parse(reasoningContent)),
+            image: [],
+            audio: [],
+            expanded: shouldExpand
+          });
+        } else {
+          chatStore.add("thinking", reasoningContent, null, null, chatId);
+        }
+      }
+      
+      // Final thinking content flush (legacy tag-based thinking)
       if (finalChat && thinkingContent.trim() && lastThinkingMessageIndex === -1) {
         chatStore.add("thinking", thinkingContent, null, null, chatId);
       }

core/http/views/backends.html

@@ -54,6 +54,11 @@
                         <span class="font-semibold text-cyan-300" x-text="installedBackends"></span>
                         <span class="text-[#94A3B8] ml-1">installed</span>
                     </a>
+                    <div class="flex items-center bg-[#101827] rounded-lg px-4 py-2 border border-[#38BDF8]/30">
+                        <i class="fas fa-microchip text-[#38BDF8] mr-2"></i>
+                        <span class="text-[#94A3B8] mr-1">Capability:</span>
+                        <span class="font-semibold text-[#38BDF8]" x-text="systemCapability"></span>
+                    </div>
                     <a href="https://localai.io/backends/" target="_blank" class="btn-primary">
                         <i class="fas fa-info-circle mr-2"></i>
                         <span>Documentation</span>
@@ -588,6 +593,7 @@ function backendsGallery() {
         totalPages: 1,
         availableBackends: 0,
         installedBackends: 0,
+        systemCapability: '',
         selectedBackend: null,
         jobProgress: {},
         notifications: [],
@@ -683,6 +689,7 @@ function backendsGallery() {
                 this.totalPages = data.totalPages || 1;
                 this.availableBackends = data.availableBackends || 0;
                 this.installedBackends = data.installedBackends || 0;
+                this.systemCapability = data.systemCapability || 'default';
             } catch (error) {
                 console.error('Error fetching backends:', error);
             } finally {

core/http/views/chat.html

@@ -41,7 +41,7 @@ SOFTWARE.
     __chatContextSize = {{ .ContextSize }};
     {{ end }}
 
-    // Store gallery configs for header icon display
+    // Store gallery configs for header icon display and model info modal
     window.__galleryConfigs = {};
     {{ $allGalleryConfigs:=.GalleryConfig }}
     {{ range $modelName, $galleryConfig := $allGalleryConfigs }}
@@ -49,6 +49,16 @@ SOFTWARE.
     {{ if $galleryConfig.Icon }}
     window.__galleryConfigs["{{$modelName}}"].Icon = "{{$galleryConfig.Icon}}";
     {{ end }}
+    {{ if $galleryConfig.Description }}
+    window.__galleryConfigs["{{$modelName}}"].Description = {{ printf "%q" $galleryConfig.Description }};
+    {{ end }}
+    {{ if $galleryConfig.URLs }}
+    window.__galleryConfigs["{{$modelName}}"].URLs = [
+      {{ range $idx, $url := $galleryConfig.URLs }}
+      {{ if $idx }},{{ end }}{{ printf "%q" $url }}
+      {{ end }}
+    ];
+    {{ end }}
     {{ end }}
 
     // Function to initialize store
@@ -326,10 +336,10 @@ SOFTWARE.
                 c += DOMPurify.sanitize(marked.parse(line));
               });
             }
-            // Set expanded state: thinking is expanded by default in non-MCP mode, collapsed in MCP mode
-            // Reasoning, tool_call, and tool_result are always collapsed by default
+            // Set expanded state: thinking and reasoning are expanded by default in non-MCP mode, collapsed in MCP mode
+            // tool_call and tool_result are always collapsed by default
             const isMCPMode = chat.mcpMode || false;
-            const shouldExpand = (role === "thinking" && !isMCPMode) || false;
+            const shouldExpand = ((role === "thinking" || role === "reasoning") && !isMCPMode) || false;
             chat.history.push({ role, content, html: c, image, audio, expanded: shouldExpand, model: messageModel });
             
             // Auto-name chat from first user message
@@ -497,6 +507,11 @@ SOFTWARE.
       activeChat.model = modelName;
       activeChat.updatedAt = Date.now();
       
+      // Update model info modal with new model
+      if (window.updateModelInfoModal) {
+        window.updateModelInfoModal(modelName);
+      }
+      
       // Get context size from data attribute
       let contextSize = null;
       if (selectedOption.dataset.contextSize) {
@@ -536,18 +551,23 @@ SOFTWARE.
       }
       
       // Update model selector to reflect the change (ensure it stays in sync)
+      // Note: We don't dispatch a change event here to avoid infinite loop
+      // The selector is already updated via user interaction or programmatic change
       const modelSelector = document.getElementById('modelSelector');
       if (modelSelector) {
         // Find and select the option matching the model
         const optionValue = 'chat/' + modelName;
         for (let i = 0; i < modelSelector.options.length; i++) {
           if (modelSelector.options[i].value === optionValue) {
-            modelSelector.selectedIndex = i;
+            // Only update if it's different to avoid unnecessary updates
+            if (modelSelector.selectedIndex !== i) {
+              modelSelector.selectedIndex = i;
+            }
             break;
           }
         }
-        // Trigger Alpine reactivity by dispatching change event
-        modelSelector.dispatchEvent(new Event('change', { bubbles: true }));
+        // Don't dispatch change event here - it would cause infinite recursion
+        // The selector is already in sync with the model
       }
       
       // Trigger MCP availability check in Alpine component
@@ -603,27 +623,52 @@ SOFTWARE.
             <div class="flex items-center justify-between gap-2">
               <label class="text-xs font-medium text-[var(--color-text-secondary)] uppercase tracking-wide flex-shrink-0">Model</label>
               <div class="flex items-center gap-1 flex-shrink-0">
-                {{ if $model }}
-                {{ $galleryConfig:= index $allGalleryConfigs $model}}
-                {{ if $galleryConfig }}
-                <button 
-                  data-twe-ripple-init 
-                  data-twe-ripple-color="light" 
-                  class="text-[var(--color-text-secondary)] hover:text-[var(--color-primary)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]" 
-                  data-modal-target="model-info-modal" 
-                  data-modal-toggle="model-info-modal"
-                  title="Model Information">
-                  <i class="fas fa-info-circle"></i>
-                </button>
-                {{ end }}
-                {{ end }}
-                {{ if $model }}
-                <a href="/models/edit/{{$model}}" 
-                   class="text-[var(--color-text-secondary)] hover:text-[var(--color-warning)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]" 
-                   title="Edit Model Configuration">
-                  <i class="fas fa-edit"></i>
-                </a>
-                {{ end }}
+                <!-- Info button - reactive to active chat model -->
+                <template x-if="$store.chat.activeChat() && $store.chat.activeChat().model && window.__galleryConfigs && window.__galleryConfigs[$store.chat.activeChat().model]">
+                  <button 
+                    data-twe-ripple-init 
+                    data-twe-ripple-color="light" 
+                    class="text-[var(--color-text-secondary)] hover:text-[var(--color-primary)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]" 
+                    data-modal-target="model-info-modal" 
+                    data-modal-toggle="model-info-modal"
+                    :data-model-name="$store.chat.activeChat().model"
+                    @click="if (window.updateModelInfoModal) { window.updateModelInfoModal($store.chat.activeChat().model, true); }"
+                    title="Model Information">
+                    <i class="fas fa-info-circle"></i>
+                  </button>
+                </template>
+                <!-- Fallback info button for initial model from server -->
+                <template x-if="(!$store.chat.activeChat() || !$store.chat.activeChat().model) && window.__galleryConfigs && window.__galleryConfigs['{{$model}}']">
+                  <button 
+                    data-twe-ripple-init 
+                    data-twe-ripple-color="light" 
+                    class="text-[var(--color-text-secondary)] hover:text-[var(--color-primary)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]" 
+                    data-modal-target="model-info-modal" 
+                    data-modal-toggle="model-info-modal"
+                    data-model-name="{{$model}}"
+                    @click="if (window.updateModelInfoModal) { window.updateModelInfoModal('{{$model}}', true); }"
+                    title="Model Information">
+                    <i class="fas fa-info-circle"></i>
+                  </button>
+                </template>
+                <!-- Edit button - reactive to active chat model -->
+                <template x-if="$store.chat.activeChat() && $store.chat.activeChat().model">
+                  <a :href="'/models/edit/' + $store.chat.activeChat().model" 
+                     class="text-[var(--color-text-secondary)] hover:text-[var(--color-warning)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]" 
+                     title="Edit Model Configuration">
+                    <i class="fas fa-edit"></i>
+                  </a>
+                </template>
+                <!-- Fallback edit button for initial model from server -->
+                <template x-if="!$store.chat.activeChat() || !$store.chat.activeChat().model">
+                  {{ if $model }}
+                  <a href="/models/edit/{{$model}}" 
+                     class="text-[var(--color-text-secondary)] hover:text-[var(--color-warning)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]" 
+                     title="Edit Model Configuration">
+                    <i class="fas fa-edit"></i>
+                  </a>
+                  {{ end }}
+                </template>
               </div>
             </div>
             <select
@@ -1488,17 +1533,14 @@ SOFTWARE.
       </div>
     </div>
 
-    <!-- Modal moved outside of sidebar to appear in center of page -->
-    {{ if $model }}
-    {{ $galleryConfig:= index $allGalleryConfigs $model}}
-    {{ if $galleryConfig }}
-    <div id="model-info-modal" tabindex="-1" aria-hidden="true" class="hidden overflow-y-auto overflow-x-hidden fixed top-0 right-0 left-0 z-50 flex justify-center items-center w-full md:inset-0 h-[calc(100%-1rem)] max-h-full">
+    <!-- Modal moved outside of sidebar to appear in center of page - Always available, content updated dynamically -->
+    <div id="model-info-modal" tabindex="-1" aria-hidden="true" class="hidden overflow-y-auto overflow-x-hidden fixed top-0 right-0 left-0 z-50 flex justify-center items-center w-full h-full md:inset-0 max-h-full" style="padding: 1rem;">
       <div class="relative p-4 w-full max-w-2xl max-h-full">
         <div class="relative p-4 w-full max-w-2xl max-h-full bg-white rounded-lg shadow dark:bg-gray-700">
           <!-- Header -->
           <div class="flex items-center justify-between p-4 md:p-5 border-b rounded-t dark:border-gray-600">
-            <h3 class="text-xl font-semibold text-gray-900 dark:text-white">{{ $model }}</h3>
-            <button class="text-gray-400 bg-transparent hover:bg-gray-200 hover:text-gray-900 rounded-lg text-sm w-8 h-8 ms-auto inline-flex justify-center items-center dark:hover:bg-gray-600 dark:hover:text-white" data-modal-hide="model-info-modal">
+            <h3 id="model-info-modal-title" class="text-xl font-semibold text-gray-900 dark:text-white">{{ if $model }}{{ $model }}{{ end }}</h3>
+            <button class="text-gray-400 bg-transparent hover:bg-gray-200 hover:text-gray-900 rounded-lg text-sm w-8 h-8 ms-auto inline-flex justify-center items-center dark:hover:bg-gray-600 dark:hover:text-white" data-modal-hide="model-info-modal" @click="if (window.closeModelInfoModal) { window.closeModelInfoModal(); }">
               <svg class="w-3 h-3" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 14 14">
                 <path stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="m1 1 6 6m0 0 6 6M7 7l6-6M7 7l-6 6"/>
               </svg>
@@ -1509,29 +1551,24 @@ SOFTWARE.
           <!-- Body -->
           <div class="p-4 md:p-5 space-y-4">
             <div class="flex justify-center items-center">
-              {{ if $galleryConfig.Icon }}<img class="lazy rounded-t-lg max-h-48 max-w-96 object-cover mt-3 entered loaded" src="{{$galleryConfig.Icon}}" loading="lazy"/>{{end}}
+              <img id="model-info-modal-icon" class="lazy rounded-t-lg max-h-48 max-w-96 object-cover mt-3 entered loaded" style="display: none;" loading="lazy"/>
             </div>
-            <div id="model-info-description" class="text-base leading-relaxed text-gray-500 dark:text-gray-400 break-words max-w-full">{{ $galleryConfig.Description }}</div>
+            <div id="model-info-description" class="text-base leading-relaxed text-gray-500 dark:text-gray-400 break-words max-w-full"></div>
             <hr>
             <p class="text-sm font-semibold text-gray-900 dark:text-white">Links</p>
-            <ul>
-              {{range $galleryConfig.URLs}}
-              <li><a href="{{ . }}" target="_blank">{{ . }}</a></li>
-              {{end}}
+            <ul id="model-info-links">
             </ul>
           </div>
 
           <!-- Footer -->
           <div class="flex items-center p-4 md:p-5 border-t border-gray-200 rounded-b dark:border-gray-600">
-            <button data-modal-hide="model-info-modal" class="py-2.5 px-5 ms-3 text-sm font-medium text-gray-900 focus:outline-none bg-white rounded-lg border border-gray-200 hover:bg-gray-100 hover:text-blue-700 focus:z-10 focus:ring-4 focus:ring-gray-100 dark:focus:ring-gray-700 dark:bg-gray-800 dark:text-gray-400 dark:border-gray-600 dark:hover:text-white dark:hover:bg-gray-700">
+            <button data-modal-hide="model-info-modal" class="py-2.5 px-5 ms-3 text-sm font-medium text-gray-900 focus:outline-none bg-white rounded-lg border border-gray-200 hover:bg-gray-100 hover:text-blue-700 focus:z-10 focus:ring-4 focus:ring-gray-100 dark:focus:ring-gray-700 dark:bg-gray-800 dark:text-gray-400 dark:border-gray-600 dark:hover:text-white dark:hover:bg-gray-700" @click="if (window.closeModelInfoModal) { window.closeModelInfoModal(); }">
               Close
             </button>
           </div>
         </div>
       </div>
     </div>
-    {{ end }}
-    {{ end }}
 
     <!-- Alpine store initialization and utilities -->
     <script>
@@ -1742,10 +1779,20 @@ SOFTWARE.
         });
 
         // Also listen for click events on modal toggle buttons
-        document.querySelectorAll('[data-modal-toggle="model-info-modal"]').forEach(button => {
-          button.addEventListener('click', () => {
+        // Use event delegation to handle dynamically created buttons
+        document.addEventListener('click', (e) => {
+          const button = e.target.closest('[data-modal-toggle="model-info-modal"]');
+          if (button) {
+            // Update modal with current model before showing
+            if (window.Alpine && window.Alpine.store("chat")) {
+              const activeChat = window.Alpine.store("chat").activeChat();
+              const modelName = activeChat ? activeChat.model : (button.dataset.modelName || (document.getElementById("chat-model") ? document.getElementById("chat-model").value : null));
+              if (modelName && window.updateModelInfoModal) {
+                window.updateModelInfoModal(modelName, true);
+              }
+            }
             setTimeout(processMarkdown, 300);
-          });
+          }
         });
 
         // Process on initial load if libraries are ready
@@ -1786,12 +1833,176 @@ SOFTWARE.
         syncModelSelectorOnLoad();
       }
 
+      // Function to update model info modal with current model
+      // Set openModal to true to actually open the modal, false to just update content
+      window.updateModelInfoModal = function(modelName, openModal = false) {
+        if (!modelName) {
+          return;
+        }
+        if (!window.__galleryConfigs) {
+          return;
+        }
+        
+        const galleryConfig = window.__galleryConfigs[modelName];
+        // Check if galleryConfig exists and has at least one property
+        if (!galleryConfig || Object.keys(galleryConfig).length === 0) {
+          // Still update the modal title even if no config, so user can see which model they clicked
+          const titleEl = document.getElementById('model-info-modal-title');
+          if (titleEl) {
+            titleEl.textContent = modelName;
+          }
+          // Show message that no info is available
+          const descEl = document.getElementById('model-info-description');
+          if (descEl) {
+            descEl.textContent = 'No additional information available for this model.';
+          }
+          const linksEl = document.getElementById('model-info-links');
+          if (linksEl) {
+            linksEl.innerHTML = '';
+          }
+          const iconEl = document.getElementById('model-info-modal-icon');
+          if (iconEl) {
+            iconEl.style.display = 'none';
+          }
+          // Only open the modal if explicitly requested
+          if (openModal) {
+            const modalElement = document.getElementById('model-info-modal');
+            if (modalElement) {
+              modalElement.classList.remove('hidden');
+              modalElement.setAttribute('aria-hidden', 'false');
+              // Add backdrop
+              let backdrop = document.querySelector('.modal-backdrop');
+              if (!backdrop) {
+                backdrop = document.createElement('div');
+                backdrop.className = 'modal-backdrop fixed inset-0 bg-gray-900 bg-opacity-50 dark:bg-opacity-80 z-40';
+                document.body.appendChild(backdrop);
+                backdrop.addEventListener('click', () => {
+                  closeModelInfoModal();
+                });
+              }
+            }
+          }
+          return;
+        }
+        
+        // Update modal title
+        const titleEl = document.getElementById('model-info-modal-title');
+        if (titleEl) {
+          titleEl.textContent = modelName;
+        }
+        
+        // Update icon
+        const iconEl = document.getElementById('model-info-modal-icon');
+        if (iconEl) {
+          if (galleryConfig.Icon) {
+            iconEl.src = galleryConfig.Icon;
+            iconEl.style.display = 'block';
+          } else {
+            iconEl.style.display = 'none';
+          }
+        }
+        
+        // Update description
+        const descEl = document.getElementById('model-info-description');
+        if (descEl) {
+          descEl.textContent = galleryConfig.Description || 'No description available.';
+        }
+        
+        // Update links
+        const linksEl = document.getElementById('model-info-links');
+        if (linksEl && galleryConfig.URLs && Array.isArray(galleryConfig.URLs) && galleryConfig.URLs.length > 0) {
+          linksEl.innerHTML = '';
+          galleryConfig.URLs.forEach(url => {
+            const li = document.createElement('li');
+            const a = document.createElement('a');
+            a.href = url;
+            a.target = '_blank';
+            a.textContent = url;
+            li.appendChild(a);
+            linksEl.appendChild(li);
+          });
+        } else if (linksEl) {
+          linksEl.innerHTML = '<li>No links available</li>';
+        }
+        
+        // Only open the modal if explicitly requested
+        if (openModal) {
+          const modalElement = document.getElementById('model-info-modal');
+          if (modalElement) {
+            // Ensure positioning classes are present (they might have been removed)
+            if (!modalElement.classList.contains('flex')) {
+              modalElement.classList.add('flex');
+            }
+            if (!modalElement.classList.contains('justify-center')) {
+              modalElement.classList.add('justify-center');
+            }
+            if (!modalElement.classList.contains('items-center')) {
+              modalElement.classList.add('items-center');
+            }
+            // Ensure fixed positioning
+            if (!modalElement.classList.contains('fixed')) {
+              modalElement.classList.add('fixed');
+            }
+            // Ensure full width and height
+            if (!modalElement.classList.contains('w-full')) {
+              modalElement.classList.add('w-full');
+            }
+            if (!modalElement.classList.contains('h-full')) {
+              modalElement.classList.add('h-full');
+            }
+            // Ensure padding is set
+            if (!modalElement.style.padding) {
+              modalElement.style.padding = '1rem';
+            }
+            // Remove hidden class if present
+            modalElement.classList.remove('hidden');
+            // Set aria-hidden to false
+            modalElement.setAttribute('aria-hidden', 'false');
+            // Add backdrop if needed
+            let backdrop = document.querySelector('.modal-backdrop');
+            if (!backdrop) {
+              backdrop = document.createElement('div');
+              backdrop.className = 'modal-backdrop fixed inset-0 bg-gray-900 bg-opacity-50 dark:bg-opacity-80 z-40';
+              document.body.appendChild(backdrop);
+              backdrop.addEventListener('click', () => {
+                window.closeModelInfoModal();
+              });
+            }
+          }
+        }
+      };
+      
+      // Function to close the model info modal
+      window.closeModelInfoModal = function() {
+        const modalElement = document.getElementById('model-info-modal');
+        if (modalElement) {
+          modalElement.classList.add('hidden');
+          modalElement.setAttribute('aria-hidden', 'true');
+        }
+        const backdrop = document.querySelector('.modal-backdrop');
+        if (backdrop) {
+          backdrop.remove();
+        }
+      };
+
       // Also sync after Alpine initializes (in case it runs after DOMContentLoaded)
+      function initializeModelInfo() {
+        syncModelSelectorOnLoad();
+        // Initialize model info modal content with current model (but don't open it)
+        if (window.updateModelInfoModal && window.Alpine && window.Alpine.store("chat")) {
+          const activeChat = window.Alpine.store("chat").activeChat();
+          const modelName = activeChat ? activeChat.model : (document.getElementById("chat-model") ? document.getElementById("chat-model").value : null);
+          if (modelName) {
+            window.updateModelInfoModal(modelName, false); // false = don't open, just update content
+          }
+        }
+      }
+      
       if (window.Alpine) {
-        Alpine.nextTick(syncModelSelectorOnLoad);
+        Alpine.nextTick(initializeModelInfo);
       } else {
         document.addEventListener('alpine:init', () => {
-          Alpine.nextTick(syncModelSelectorOnLoad);
+          Alpine.nextTick(initializeModelInfo);
         });
       }
     </script>

core/schema/anthropic.go

@@ -0,0 +1,176 @@
+package schema
+
+import (
+	"context"
+	"encoding/json"
+)
+
+// AnthropicRequest represents a request to the Anthropic Messages API
+// https://docs.anthropic.com/claude/reference/messages_post
+type AnthropicRequest struct {
+	Model         string             `json:"model"`
+	Messages      []AnthropicMessage `json:"messages"`
+	MaxTokens     int                `json:"max_tokens"`
+	Metadata      map[string]string  `json:"metadata,omitempty"`
+	StopSequences []string           `json:"stop_sequences,omitempty"`
+	Stream        bool               `json:"stream,omitempty"`
+	System        string             `json:"system,omitempty"`
+	Temperature   *float64           `json:"temperature,omitempty"`
+	TopK          *int               `json:"top_k,omitempty"`
+	TopP          *float64           `json:"top_p,omitempty"`
+	Tools         []AnthropicTool    `json:"tools,omitempty"`
+	ToolChoice    interface{}        `json:"tool_choice,omitempty"`
+
+	// Internal fields for request handling
+	Context context.Context    `json:"-"`
+	Cancel  context.CancelFunc `json:"-"`
+}
+
+// ModelName implements the LocalAIRequest interface
+func (ar *AnthropicRequest) ModelName(s *string) string {
+	if s != nil {
+		ar.Model = *s
+	}
+	return ar.Model
+}
+
+// AnthropicTool represents a tool definition in the Anthropic format
+type AnthropicTool struct {
+	Name        string                 `json:"name"`
+	Description string                 `json:"description,omitempty"`
+	InputSchema map[string]interface{} `json:"input_schema"`
+}
+
+// AnthropicMessage represents a message in the Anthropic format
+type AnthropicMessage struct {
+	Role    string      `json:"role"`
+	Content interface{} `json:"content"`
+}
+
+// AnthropicContentBlock represents a content block in an Anthropic message
+type AnthropicContentBlock struct {
+	Type       string                 `json:"type"`
+	Text       string                 `json:"text,omitempty"`
+	Source     *AnthropicImageSource  `json:"source,omitempty"`
+	ID         string                 `json:"id,omitempty"`
+	Name       string                 `json:"name,omitempty"`
+	Input      map[string]interface{} `json:"input,omitempty"`
+	ToolUseID  string                 `json:"tool_use_id,omitempty"`
+	Content    interface{}            `json:"content,omitempty"`
+	IsError    *bool                  `json:"is_error,omitempty"`
+}
+
+// AnthropicImageSource represents an image source in Anthropic format
+type AnthropicImageSource struct {
+	Type      string `json:"type"`
+	MediaType string `json:"media_type"`
+	Data      string `json:"data"`
+}
+
+// AnthropicResponse represents a response from the Anthropic Messages API
+type AnthropicResponse struct {
+	ID           string                  `json:"id"`
+	Type         string                  `json:"type"`
+	Role         string                  `json:"role"`
+	Content      []AnthropicContentBlock `json:"content"`
+	Model        string                  `json:"model"`
+	StopReason   *string                 `json:"stop_reason"`
+	StopSequence *string                 `json:"stop_sequence,omitempty"`
+	Usage        AnthropicUsage          `json:"usage"`
+}
+
+// AnthropicUsage represents token usage in Anthropic format
+type AnthropicUsage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+}
+
+// AnthropicStreamEvent represents a streaming event from the Anthropic API
+type AnthropicStreamEvent struct {
+	Type         string                  `json:"type"`
+	Index        int                     `json:"index,omitempty"`
+	ContentBlock *AnthropicContentBlock  `json:"content_block,omitempty"`
+	Delta        *AnthropicStreamDelta   `json:"delta,omitempty"`
+	Message      *AnthropicStreamMessage `json:"message,omitempty"`
+	Usage        *AnthropicUsage         `json:"usage,omitempty"`
+}
+
+// AnthropicStreamDelta represents the delta in a streaming response
+type AnthropicStreamDelta struct {
+	Type         string  `json:"type,omitempty"`
+	Text         string  `json:"text,omitempty"`
+	PartialJSON  string  `json:"partial_json,omitempty"`
+	StopReason   *string `json:"stop_reason,omitempty"`
+	StopSequence *string `json:"stop_sequence,omitempty"`
+}
+
+// AnthropicStreamMessage represents the message object in streaming events
+type AnthropicStreamMessage struct {
+	ID           string                  `json:"id"`
+	Type         string                  `json:"type"`
+	Role         string                  `json:"role"`
+	Content      []AnthropicContentBlock `json:"content"`
+	Model        string                  `json:"model"`
+	StopReason   *string                 `json:"stop_reason"`
+	StopSequence *string                 `json:"stop_sequence,omitempty"`
+	Usage        AnthropicUsage          `json:"usage"`
+}
+
+// AnthropicErrorResponse represents an error response from the Anthropic API
+type AnthropicErrorResponse struct {
+	Type  string         `json:"type"`
+	Error AnthropicError `json:"error"`
+}
+
+// AnthropicError represents an error in the Anthropic format
+type AnthropicError struct {
+	Type    string `json:"type"`
+	Message string `json:"message"`
+}
+
+// GetStringContent extracts the string content from an AnthropicMessage
+// Content can be either a string or an array of content blocks
+func (m *AnthropicMessage) GetStringContent() string {
+	switch content := m.Content.(type) {
+	case string:
+		return content
+	case []interface{}:
+		var result string
+		for _, block := range content {
+			if blockMap, ok := block.(map[string]interface{}); ok {
+				if blockMap["type"] == "text" {
+					if text, ok := blockMap["text"].(string); ok {
+						result += text
+					}
+				}
+			}
+		}
+		return result
+	}
+	return ""
+}
+
+// GetContentBlocks extracts content blocks from an AnthropicMessage
+func (m *AnthropicMessage) GetContentBlocks() []AnthropicContentBlock {
+	switch content := m.Content.(type) {
+	case string:
+		return []AnthropicContentBlock{{Type: "text", Text: content}}
+	case []interface{}:
+		var blocks []AnthropicContentBlock
+		for _, block := range content {
+			if blockMap, ok := block.(map[string]interface{}); ok {
+				cb := AnthropicContentBlock{}
+				data, err := json.Marshal(blockMap)
+				if err != nil {
+					continue
+				}
+				if err := json.Unmarshal(data, &cb); err != nil {
+					continue
+				}
+				blocks = append(blocks, cb)
+			}
+		}
+		return blocks
+	}
+	return nil
+}

core/schema/anthropic_test.go

@@ -0,0 +1,216 @@
+package schema_test
+
+import (
+	"encoding/json"
+
+	"github.com/mudler/LocalAI/core/schema"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("Anthropic Schema", func() {
+	Describe("AnthropicRequest", func() {
+		It("should unmarshal a valid request", func() {
+			jsonData := `{
+				"model": "claude-3-sonnet-20240229",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello, world!"}
+				],
+				"system": "You are a helpful assistant.",
+				"temperature": 0.7
+			}`
+
+			var req schema.AnthropicRequest
+			err := json.Unmarshal([]byte(jsonData), &req)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(req.Model).To(Equal("claude-3-sonnet-20240229"))
+			Expect(req.MaxTokens).To(Equal(1024))
+			Expect(len(req.Messages)).To(Equal(1))
+			Expect(req.System).To(Equal("You are a helpful assistant."))
+			Expect(*req.Temperature).To(Equal(0.7))
+		})
+
+		It("should unmarshal a request with tools", func() {
+			jsonData := `{
+				"model": "claude-3-sonnet-20240229",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"}
+				],
+				"tools": [
+					{
+						"name": "get_weather",
+						"description": "Get the current weather",
+						"input_schema": {
+							"type": "object",
+							"properties": {
+								"location": {"type": "string"}
+							}
+						}
+					}
+				],
+				"tool_choice": {"type": "tool", "name": "get_weather"}
+			}`
+
+			var req schema.AnthropicRequest
+			err := json.Unmarshal([]byte(jsonData), &req)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(len(req.Tools)).To(Equal(1))
+			Expect(req.Tools[0].Name).To(Equal("get_weather"))
+			Expect(req.Tools[0].Description).To(Equal("Get the current weather"))
+			Expect(req.ToolChoice).ToNot(BeNil())
+		})
+
+		It("should implement LocalAIRequest interface", func() {
+			req := &schema.AnthropicRequest{Model: "test-model"}
+			Expect(req.ModelName(nil)).To(Equal("test-model"))
+
+			newModel := "new-model"
+			Expect(req.ModelName(&newModel)).To(Equal("new-model"))
+			Expect(req.Model).To(Equal("new-model"))
+		})
+	})
+
+	Describe("AnthropicMessage", func() {
+		It("should get string content from string content", func() {
+			msg := schema.AnthropicMessage{
+				Role:    "user",
+				Content: "Hello, world!",
+			}
+			Expect(msg.GetStringContent()).To(Equal("Hello, world!"))
+		})
+
+		It("should get string content from array content", func() {
+			msg := schema.AnthropicMessage{
+				Role: "user",
+				Content: []interface{}{
+					map[string]interface{}{"type": "text", "text": "Hello, "},
+					map[string]interface{}{"type": "text", "text": "world!"},
+				},
+			}
+			Expect(msg.GetStringContent()).To(Equal("Hello, world!"))
+		})
+
+		It("should get content blocks from string content", func() {
+			msg := schema.AnthropicMessage{
+				Role:    "user",
+				Content: "Hello, world!",
+			}
+			blocks := msg.GetContentBlocks()
+			Expect(len(blocks)).To(Equal(1))
+			Expect(blocks[0].Type).To(Equal("text"))
+			Expect(blocks[0].Text).To(Equal("Hello, world!"))
+		})
+
+		It("should get content blocks from array content", func() {
+			msg := schema.AnthropicMessage{
+				Role: "user",
+				Content: []interface{}{
+					map[string]interface{}{"type": "text", "text": "Hello"},
+					map[string]interface{}{"type": "image", "source": map[string]interface{}{"type": "base64", "data": "abc123"}},
+				},
+			}
+			blocks := msg.GetContentBlocks()
+			Expect(len(blocks)).To(Equal(2))
+			Expect(blocks[0].Type).To(Equal("text"))
+			Expect(blocks[0].Text).To(Equal("Hello"))
+		})
+	})
+
+	Describe("AnthropicResponse", func() {
+		It("should marshal a valid response", func() {
+			stopReason := "end_turn"
+			resp := schema.AnthropicResponse{
+				ID:         "msg_123",
+				Type:       "message",
+				Role:       "assistant",
+				Model:      "claude-3-sonnet-20240229",
+				StopReason: &stopReason,
+				Content: []schema.AnthropicContentBlock{
+					{Type: "text", Text: "Hello!"},
+				},
+				Usage: schema.AnthropicUsage{
+					InputTokens:  10,
+					OutputTokens: 5,
+				},
+			}
+
+			data, err := json.Marshal(resp)
+			Expect(err).ToNot(HaveOccurred())
+
+			var result map[string]interface{}
+			err = json.Unmarshal(data, &result)
+			Expect(err).ToNot(HaveOccurred())
+
+			Expect(result["id"]).To(Equal("msg_123"))
+			Expect(result["type"]).To(Equal("message"))
+			Expect(result["role"]).To(Equal("assistant"))
+			Expect(result["stop_reason"]).To(Equal("end_turn"))
+		})
+
+		It("should marshal a response with tool use", func() {
+			stopReason := "tool_use"
+			resp := schema.AnthropicResponse{
+				ID:         "msg_123",
+				Type:       "message",
+				Role:       "assistant",
+				Model:      "claude-3-sonnet-20240229",
+				StopReason: &stopReason,
+				Content: []schema.AnthropicContentBlock{
+					{
+						Type: "tool_use",
+						ID:   "toolu_123",
+						Name: "get_weather",
+						Input: map[string]interface{}{
+							"location": "San Francisco",
+						},
+					},
+				},
+				Usage: schema.AnthropicUsage{
+					InputTokens:  10,
+					OutputTokens: 5,
+				},
+			}
+
+			data, err := json.Marshal(resp)
+			Expect(err).ToNot(HaveOccurred())
+
+			var result map[string]interface{}
+			err = json.Unmarshal(data, &result)
+			Expect(err).ToNot(HaveOccurred())
+
+			Expect(result["stop_reason"]).To(Equal("tool_use"))
+			content := result["content"].([]interface{})
+			Expect(len(content)).To(Equal(1))
+			toolUse := content[0].(map[string]interface{})
+			Expect(toolUse["type"]).To(Equal("tool_use"))
+			Expect(toolUse["id"]).To(Equal("toolu_123"))
+			Expect(toolUse["name"]).To(Equal("get_weather"))
+		})
+	})
+
+	Describe("AnthropicErrorResponse", func() {
+		It("should marshal an error response", func() {
+			resp := schema.AnthropicErrorResponse{
+				Type: "error",
+				Error: schema.AnthropicError{
+					Type:    "invalid_request_error",
+					Message: "max_tokens is required",
+				},
+			}
+
+			data, err := json.Marshal(resp)
+			Expect(err).ToNot(HaveOccurred())
+
+			var result map[string]interface{}
+			err = json.Unmarshal(data, &result)
+			Expect(err).ToNot(HaveOccurred())
+
+			Expect(result["type"]).To(Equal("error"))
+			errorObj := result["error"].(map[string]interface{})
+			Expect(errorObj["type"]).To(Equal("invalid_request_error"))
+			Expect(errorObj["message"]).To(Equal("max_tokens is required"))
+		})
+	})
+})

core/schema/message.go

@@ -27,6 +27,9 @@ type Message struct {
 	FunctionCall interface{} `json:"function_call,omitempty" yaml:"function_call,omitempty"`
 
 	ToolCalls []ToolCall `json:"tool_calls,omitempty" yaml:"tool_call,omitempty"`
+
+	// Reasoning content extracted from <thinking>...</thinking> tags
+	Reasoning *string `json:"reasoning,omitempty" yaml:"reasoning,omitempty"`
 }
 
 type ToolCall struct {
@@ -78,8 +81,8 @@ func (messages Messages) ToProto() []*proto.Message {
 			}
 		}
 
-		// Note: tool_call_id and reasoning_content are not in schema.Message yet
-		// They may need to be added to schema.Message if needed in the future
+		// Note: tool_call_id is not in schema.Message yet
+		// Reasoning field is now available in schema.Message but not yet in proto.Message
 	}
 	return protoMessages
 }

gallery/index.yaml

@@ -1,4 +1,56 @@
 ---
+- name: "qwen3-vl-reranker-8b"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/mradermacher/Qwen3-VL-Reranker-8B-GGUF
+  description: |
+    **Model Name:** Qwen3-VL-Reranker-8B
+    **Base Model:** Qwen/Qwen3-VL-Reranker-8B
+
+    **Description:**
+    A high-performance multimodal reranking model for state-of-the-art cross-modal search. It supports 30+ languages and handles text, images, screenshots, videos, and mixed modalities. With 8B parameters and a 32K context length, it refines retrieval results by combining embedding vectors with precise relevance scores. Optimized for efficiency, it supports quantized versions (e.g., Q8_0, Q4_K_M) and is ideal for applications requiring accurate multimodal content matching.
+
+    **Key Features:**
+      - **Multimodal**: Text, images, videos, and mixed content.
+      - **Language Support**: 30+ languages.
+      - **Quantization**: Available in Q8_0 (best quality), Q4_K_M (fast, recommended), and lower-precision options.
+      - **Performance**: Outperforms base models in retrieval tasks (e.g., JinaVDR, ViDoRe v3).
+      - **Use Case**: Enhances search pipelines by refining embeddings with precise relevance scores.
+
+    **Downloads:**
+      - [GGUF Files](https://huggingface.co/mradermacher/Qwen3-VL-Reranker-8B-GGUF) (e.g., `Qwen3-VL-Reranker-8B.Q8_0.gguf`).
+
+    **Usage:**
+      - Requires `transformers`, `qwen-vl-utils`, and `torch`.
+      - Example: `from scripts.qwen3_vl_reranker import Qwen3VLReranker; model = Qwen3VLReranker(...)`
+
+    **Citation:**
+    @article{qwen3vlembedding, ...}
+
+    This description emphasizes its capabilities, efficiency, and versatility for multimodal search tasks.
+  overrides:
+    parameters:
+      model: llama-cpp/models/Qwen3-VL-Reranker-8B.Q4_K_M.gguf
+    name: Qwen3-VL-Reranker-8B-GGUF
+    backend: llama-cpp
+    template:
+      use_tokenizer_template: true
+    known_usecases:
+      - chat
+    function:
+      grammar:
+        disable: true
+    mmproj: llama-cpp/mmproj/Qwen3-VL-Reranker-8B.mmproj-f16.gguf
+    description: Imported from https://huggingface.co/mradermacher/Qwen3-VL-Reranker-8B-GGUF
+    options:
+      - use_jinja:true
+  files:
+    - filename: llama-cpp/models/Qwen3-VL-Reranker-8B.Q4_K_M.gguf
+      sha256: f73e62ea68abf741c3e713af823cfb4d2fd2ca35c8b68277b87b4b3d8570b66d
+      uri: https://huggingface.co/mradermacher/Qwen3-VL-Reranker-8B-GGUF/resolve/main/Qwen3-VL-Reranker-8B.Q4_K_M.gguf
+    - filename: llama-cpp/mmproj/Qwen3-VL-Reranker-8B.mmproj-f16.gguf
+      sha256: 15cd9bd4882dae771344f0ac204fce07de91b47c1438ada3861dfc817403c31e
+      uri: https://huggingface.co/mradermacher/Qwen3-VL-Reranker-8B-GGUF/resolve/main/Qwen3-VL-Reranker-8B.mmproj-f16.gguf
 - name: "liquidai.lfm2-2.6b-transcript"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:

go.mod

@@ -9,6 +9,7 @@ require (
 	fyne.io/fyne/v2 v2.7.1
 	github.com/Masterminds/sprig/v3 v3.3.0
 	github.com/alecthomas/kong v1.13.0
+	github.com/anthropics/anthropic-sdk-go v1.19.0
 	github.com/charmbracelet/glamour v0.10.0
 	github.com/containerd/containerd v1.7.30
 	github.com/ebitengine/purego v0.9.1
@@ -58,6 +59,7 @@ require (
 	go.opentelemetry.io/otel/metric v1.39.0
 	go.opentelemetry.io/otel/sdk/metric v1.39.0
 	google.golang.org/grpc v1.78.0
+	google.golang.org/protobuf v1.36.10
 	gopkg.in/yaml.v2 v2.4.0
 	gopkg.in/yaml.v3 v3.0.1
 	oras.land/oras-go/v2 v2.6.0
@@ -67,8 +69,11 @@ require (
 	github.com/ghodss/yaml v1.0.0 // indirect
 	github.com/labstack/gommon v0.4.2 // indirect
 	github.com/swaggo/files/v2 v2.0.2 // indirect
+	github.com/tidwall/gjson v1.18.0 // indirect
+	github.com/tidwall/match v1.1.1 // indirect
+	github.com/tidwall/pretty v1.2.1 // indirect
+	github.com/tidwall/sjson v1.2.5 // indirect
 	github.com/valyala/fasttemplate v1.2.2 // indirect
-	google.golang.org/protobuf v1.36.10 // indirect
 )
 
 require (

go.sum

@@ -44,6 +44,8 @@ github.com/andybalholm/brotli v1.0.1/go.mod h1:loMXtMfwqflxFJPmdbJO0a3KNoPuLBgiu
 github.com/andybalholm/brotli v1.2.0 h1:ukwgCxwYrmACq68yiUqwIWnGY0cTPox/M94sVwToPjQ=
 github.com/andybalholm/brotli v1.2.0/go.mod h1:rzTDkvFWvIrjDXZHkuS16NPggd91W3kUSvPlQ1pLaKY=
 github.com/anmitsu/go-shlex v0.0.0-20161002113705-648efa622239/go.mod h1:2FmKhYUyUczH0OGQWaF5ceTx0UBShxjsH6f8oGKYe2c=
+github.com/anthropics/anthropic-sdk-go v1.19.0 h1:mO6E+ffSzLRvR/YUH9KJC0uGw0uV8GjISIuzem//3KE=
+github.com/anthropics/anthropic-sdk-go v1.19.0/go.mod h1:WTz31rIUHUHqai2UslPpw5CwXrQP3geYBioRV4WOLvE=
 github.com/aymanbagabas/go-osc52/v2 v2.0.1 h1:HwpRHbFMcZLEVr42D4p7XBqjyuxQH5SMiErDT4WkJ2k=
 github.com/aymanbagabas/go-osc52/v2 v2.0.1/go.mod h1:uYgXzlJ7ZpABp8OJ+exZzJJhRNQ2ASbcXHWsFqH8hp8=
 github.com/aymanbagabas/go-udiff v0.2.0 h1:TK0fH4MteXUDspT88n8CKzvK0X9O2xu9yQjWpi6yML8=
@@ -762,10 +764,12 @@ github.com/swaggo/swag v1.16.6/go.mod h1:ngP2etMK5a0P3QBizic5MEwpRmluJZPHjXcMoj4
 github.com/tarm/serial v0.0.0-20180830185346-98f6abe2eb07/go.mod h1:kDXzergiv9cbyO7IOYJZWg1U88JhDg3PB6klq9Hg2pA=
 github.com/testcontainers/testcontainers-go v0.40.0 h1:pSdJYLOVgLE8YdUY2FHQ1Fxu+aMnb6JfVz1mxk7OeMU=
 github.com/testcontainers/testcontainers-go v0.40.0/go.mod h1:FSXV5KQtX2HAMlm7U3APNyLkkap35zNLxukw9oBi/MY=
+github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
 github.com/tidwall/gjson v1.18.0 h1:FIDeeyB800efLX89e5a8Y0BNH+LOngJyGrIWxG2FKQY=
 github.com/tidwall/gjson v1.18.0/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
 github.com/tidwall/match v1.1.1 h1:+Ho715JplO36QYgwN9PGYNhgZvoUSc9X2c80KVTi+GA=
 github.com/tidwall/match v1.1.1/go.mod h1:eRSPERbgtNPcGhD8UCthc6PmLEQXEWd3PRB5JTxsfmM=
+github.com/tidwall/pretty v1.2.0/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
 github.com/tidwall/pretty v1.2.1 h1:qjsOFOWWQl+N3RsoF5/ssm1pHmJJwhjlSbZ51I6wMl4=
 github.com/tidwall/pretty v1.2.1/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
 github.com/tidwall/sjson v1.2.5 h1:kLy8mja+1c9jlljvWTlSazM7cKDRfJuR/bOJhcY5NcY=

pkg/functions/reasoning.go

@@ -0,0 +1,114 @@
+package functions
+
+import (
+	"strings"
+)
+
+// ExtractReasoning extracts reasoning content from thinking tags and returns
+// both the extracted reasoning and the cleaned content (with tags removed).
+// It handles <thinking>...</thinking> and <think>...</think> tags.
+// Multiple reasoning blocks are concatenated with newlines.
+func ExtractReasoning(content string) (reasoning string, cleanedContent string) {
+	if content == "" {
+		return "", content
+	}
+
+	var reasoningParts []string
+	var cleanedParts []string
+	remaining := content
+
+	// Define tag pairs to look for
+	tagPairs := []struct {
+		start string
+		end   string
+	}{
+		{"<thinking>", "</thinking>"},
+		{"<think>", "</think>"},
+	}
+
+	// Track the last position we've processed
+	lastPos := 0
+
+	for {
+		// Find the earliest tag start
+		earliestStart := -1
+		earliestEnd := -1
+		isUnclosed := false
+		var matchedTag struct {
+			start string
+			end   string
+		}
+
+		for _, tagPair := range tagPairs {
+			startIdx := strings.Index(remaining[lastPos:], tagPair.start)
+			if startIdx == -1 {
+				continue
+			}
+			startIdx += lastPos
+
+			// Find the corresponding end tag
+			endIdx := strings.Index(remaining[startIdx+len(tagPair.start):], tagPair.end)
+			if endIdx == -1 {
+				// Unclosed tag - extract what we have
+				if earliestStart == -1 || startIdx < earliestStart {
+					earliestStart = startIdx
+					earliestEnd = len(remaining)
+					isUnclosed = true
+					matchedTag = tagPair
+				}
+				continue
+			}
+			endIdx += startIdx + len(tagPair.start)
+
+			// Found a complete tag pair
+			if earliestStart == -1 || startIdx < earliestStart {
+				earliestStart = startIdx
+				earliestEnd = endIdx + len(tagPair.end)
+				isUnclosed = false
+				matchedTag = tagPair
+			}
+		}
+
+		if earliestStart == -1 {
+			// No more tags found, add remaining content
+			if lastPos < len(remaining) {
+				cleanedParts = append(cleanedParts, remaining[lastPos:])
+			}
+			break
+		}
+
+		// Add content before the tag
+		if earliestStart > lastPos {
+			cleanedParts = append(cleanedParts, remaining[lastPos:earliestStart])
+		}
+
+		// Extract reasoning content
+		reasoningStart := earliestStart + len(matchedTag.start)
+		// For unclosed tags, earliestEnd is already at the end of the string
+		// For closed tags, earliestEnd points to after the closing tag, so we subtract the end tag length
+		var reasoningEnd int
+		if isUnclosed {
+			// Unclosed tag - extract everything to the end
+			reasoningEnd = len(remaining)
+		} else {
+			// Closed tag - exclude the end tag
+			reasoningEnd = earliestEnd - len(matchedTag.end)
+		}
+		if reasoningEnd > reasoningStart {
+			reasoningContent := strings.TrimSpace(remaining[reasoningStart:reasoningEnd])
+			if reasoningContent != "" {
+				reasoningParts = append(reasoningParts, reasoningContent)
+			}
+		}
+
+		// Move past this tag
+		lastPos = earliestEnd
+	}
+
+	// Combine reasoning parts
+	reasoning = strings.Join(reasoningParts, "\n\n")
+	// Combine cleaned content parts
+	cleanedContent = strings.Join(cleanedParts, "")
+
+	return reasoning, cleanedContent
+}

pkg/functions/reasoning_test.go

@@ -0,0 +1,261 @@
+package functions_test
+
+import (
+	"strings"
+
+	. "github.com/mudler/LocalAI/pkg/functions"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("ExtractReasoning", func() {
+	Context("when content has no reasoning tags", func() {
+		It("should return empty reasoning and original content", func() {
+			content := "This is regular content without any tags."
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal(content))
+		})
+
+		It("should handle empty string", func() {
+			content := ""
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(BeEmpty())
+		})
+
+		It("should handle content with only whitespace", func() {
+			content := "   \n\t  "
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal(content))
+		})
+	})
+
+	Context("when content has <thinking> tags", func() {
+		It("should extract reasoning from single thinking block", func() {
+			content := "Some text <thinking>This is my reasoning</thinking> More text"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("This is my reasoning"))
+			Expect(cleaned).To(Equal("Some text  More text"))
+		})
+
+		It("should extract reasoning and preserve surrounding content", func() {
+			content := "Before <thinking>Reasoning here</thinking> After"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Reasoning here"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle thinking block at the start", func() {
+			content := "<thinking>Start reasoning</thinking> Regular content"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Start reasoning"))
+			Expect(cleaned).To(Equal(" Regular content"))
+		})
+
+		It("should handle thinking block at the end", func() {
+			content := "Regular content <thinking>End reasoning</thinking>"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("End reasoning"))
+			Expect(cleaned).To(Equal("Regular content "))
+		})
+
+		It("should handle only thinking block", func() {
+			content := "<thinking>Only reasoning</thinking>"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Only reasoning"))
+			Expect(cleaned).To(BeEmpty())
+		})
+
+		It("should trim whitespace from reasoning content", func() {
+			content := "Text <thinking>  \n  Reasoning with spaces  \n  </thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Reasoning with spaces"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has <think> tags", func() {
+		It("should extract reasoning from redacted_reasoning block", func() {
+			content := "Text <think>Redacted reasoning</think> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Redacted reasoning"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should handle redacted_reasoning with multiline content", func() {
+			content := "Before <think>Line 1\nLine 2\nLine 3</think> After"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle redacted_reasoning with complex content", func() {
+			content := "Start <think>Complex reasoning\nwith\nmultiple\nlines</think> End"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Complex reasoning\nwith\nmultiple\nlines"))
+			Expect(cleaned).To(Equal("Start  End"))
+		})
+	})
+
+	Context("when content has multiple reasoning blocks", func() {
+		It("should concatenate multiple thinking blocks with newlines", func() {
+			content := "Text <thinking>First</thinking> Middle <thinking>Second</thinking> End"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("First\n\nSecond"))
+			Expect(cleaned).To(Equal("Text  Middle  End"))
+		})
+
+		It("should handle multiple different tag types", func() {
+			content := "A <thinking>One</thinking> B <think>Two</think> C <think>Three</think> D"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(ContainSubstring("One"))
+			Expect(reasoning).To(ContainSubstring("Two"))
+			Expect(reasoning).To(ContainSubstring("Three"))
+			Expect(cleaned).To(Equal("A  B  C  D"))
+		})
+
+		It("should handle nested tags correctly (extracts first match)", func() {
+			content := "Text <thinking>Outer <think>Inner</think></thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			// Should extract the outer thinking block
+			Expect(reasoning).To(ContainSubstring("Outer"))
+			Expect(reasoning).To(ContainSubstring("Inner"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has unclosed reasoning tags", func() {
+		It("should extract unclosed thinking block", func() {
+			content := "Text <thinking>Unclosed reasoning"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Unclosed reasoning"))
+			Expect(cleaned).To(Equal("Text "))
+		})
+
+		It("should extract unclosed think block", func() {
+			content := "Before <think>Incomplete"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Incomplete"))
+			Expect(cleaned).To(Equal("Before "))
+		})
+
+		It("should extract unclosed redacted_reasoning block", func() {
+			content := "Start <think>Partial reasoning content"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Partial reasoning content"))
+			Expect(cleaned).To(Equal("Start "))
+		})
+
+		It("should handle unclosed tag at the end", func() {
+			content := "Regular content <thinking>Unclosed at end"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Unclosed at end"))
+			Expect(cleaned).To(Equal("Regular content "))
+		})
+	})
+
+	Context("when content has empty reasoning blocks", func() {
+		It("should ignore empty thinking block", func() {
+			content := "Text <thinking></thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should ignore thinking block with only whitespace", func() {
+			content := "Text <thinking>   \n\t  </thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has reasoning tags with special characters", func() {
+		It("should handle reasoning with newlines", func() {
+			content := "Before <thinking>Line 1\nLine 2\nLine 3</thinking> After"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle reasoning with code blocks", func() {
+			content := "Text <thinking>Reasoning with ```code``` blocks</thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Reasoning with ```code``` blocks"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should handle reasoning with JSON", func() {
+			content := "Before <think>{\"key\": \"value\"}</think> After"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("{\"key\": \"value\"}"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle reasoning with HTML-like content", func() {
+			content := "Text <thinking>Reasoning with <tags> inside</thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Reasoning with <tags> inside"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has reasoning mixed with regular content", func() {
+		It("should preserve content order correctly", func() {
+			content := "Start <thinking>Reasoning</thinking> Middle <think>More reasoning</think> End"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(ContainSubstring("Reasoning"))
+			Expect(reasoning).To(ContainSubstring("More reasoning"))
+			Expect(cleaned).To(Equal("Start  Middle  End"))
+		})
+
+		It("should handle reasoning in the middle of a sentence", func() {
+			content := "This is a <thinking>reasoning</thinking> sentence."
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("reasoning"))
+			Expect(cleaned).To(Equal("This is a  sentence."))
+		})
+	})
+
+	Context("edge cases", func() {
+		It("should handle content with only opening tag", func() {
+			content := "<thinking>"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal(""))
+		})
+
+		It("should handle content with only closing tag", func() {
+			content := "</thinking>"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal("</thinking>"))
+		})
+
+		It("should handle mismatched tags", func() {
+			content := "<thinking>Content</think>"
+			reasoning, cleaned := ExtractReasoning(content)
+			// Should extract unclosed thinking block
+			Expect(reasoning).To(ContainSubstring("Content"))
+			Expect(cleaned).To(Equal(""))
+		})
+
+		It("should handle very long reasoning content", func() {
+			longReasoning := strings.Repeat("This is reasoning content. ", 100)
+			content := "Text <thinking>" + longReasoning + "</thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			// TrimSpace is applied, so we need to account for that
+			Expect(reasoning).To(Equal(strings.TrimSpace(longReasoning)))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should handle reasoning with unicode characters", func() {
+			content := "Text <thinking>Reasoning with 中文 and emoji 🧠</thinking> More"
+			reasoning, cleaned := ExtractReasoning(content)
+			Expect(reasoning).To(Equal("Reasoning with 中文 and emoji 🧠"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+})

pkg/model/process.go

@@ -24,8 +24,6 @@ func (ml *ModelLoader) deleteProcess(s string) error {
 		return fmt.Errorf("model %s not found", s)
 	}
 
-	defer delete(ml.models, s)
-
 	retries := 1
 	for model.GRPC(false, ml.wd).IsBusy() {
 		xlog.Debug("Model busy. Waiting.", "model", s)
@@ -48,6 +46,7 @@ func (ml *ModelLoader) deleteProcess(s string) error {
 	if process == nil {
 		xlog.Error("No process", "model", s)
 		// Nothing to do as there is no process
+		delete(ml.models, s)
 		return nil
 	}
 
@@ -56,6 +55,10 @@ func (ml *ModelLoader) deleteProcess(s string) error {
 		xlog.Error("(deleteProcess) error while deleting process", "error", err, "model", s)
 	}
 
+	if err == nil {
+		delete(ml.models, s)
+	}
+
 	return err
 }
 

pkg/system/capabilities.go

@@ -12,15 +12,17 @@ import (
 )
 
 const (
+	// Public constants - used by tests and external packages
+	Nvidia = "nvidia"
+	AMD    = "amd"
+	Intel  = "intel"
+
+	// Private constants - only used within this package
 	defaultCapability = "default"
 	nvidiaL4T         = "nvidia-l4t"
 	darwinX86         = "darwin-x86"
 	metal             = "metal"
-	nvidia            = "nvidia"
-
-	amd    = "amd"
-	intel  = "intel"
-	vulkan = "vulkan"
+	vulkan            = "vulkan"
 
 	nvidiaCuda13    = "nvidia-cuda-13"
 	nvidiaCuda12    = "nvidia-cuda-12"
@@ -30,6 +32,16 @@ const (
 	capabilityEnv        = "LOCALAI_FORCE_META_BACKEND_CAPABILITY"
 	capabilityRunFileEnv = "LOCALAI_FORCE_META_BACKEND_CAPABILITY_RUN_FILE"
 	defaultRunFile       = "/run/localai/capability"
+
+	// Backend detection tokens (private)
+	backendTokenDarwin = "darwin"
+	backendTokenMLX    = "mlx"
+	backendTokenMetal  = "metal"
+	backendTokenL4T    = "l4t"
+	backendTokenCUDA   = "cuda"
+	backendTokenROCM   = "rocm"
+	backendTokenHIP    = "hip"
+	backendTokenSYCL   = "sycl"
 )
 
 var (
@@ -96,7 +108,7 @@ func (s *SystemState) getSystemCapabilities() string {
 
 	// If arm64 on linux and a nvidia gpu is detected, we will return nvidia-l4t
 	if runtime.GOOS == "linux" && runtime.GOARCH == "arm64" {
-		if s.GPUVendor == nvidia {
+		if s.GPUVendor == Nvidia {
 			xlog.Info("Using nvidia-l4t capability (arm64 on linux)", "env", capabilityEnv)
 			if cuda13DirExists {
 				return nvidiaL4TCuda13
@@ -131,7 +143,6 @@ func (s *SystemState) getSystemCapabilities() string {
 	return s.GPUVendor
 }
 
-
 // BackendPreferenceTokens returns a list of substrings that represent the preferred
 // backend implementation order for the current system capability. Callers can use
 // these tokens to select the most appropriate concrete backend among multiple
@@ -139,19 +150,76 @@ func (s *SystemState) getSystemCapabilities() string {
 func (s *SystemState) BackendPreferenceTokens() []string {
 	capStr := strings.ToLower(s.getSystemCapabilities())
 	switch {
-	case strings.HasPrefix(capStr, nvidia):
-		return []string{"cuda", "vulkan", "cpu"}
-	case strings.HasPrefix(capStr, amd):
-		return []string{"rocm", "hip", "vulkan", "cpu"}
-	case strings.HasPrefix(capStr, intel):
-		return []string{"sycl", intel, "cpu"}
+	case strings.HasPrefix(capStr, Nvidia):
+		return []string{backendTokenCUDA, vulkan, "cpu"}
+	case strings.HasPrefix(capStr, AMD):
+		return []string{backendTokenROCM, backendTokenHIP, vulkan, "cpu"}
+	case strings.HasPrefix(capStr, Intel):
+		return []string{backendTokenSYCL, Intel, "cpu"}
 	case strings.HasPrefix(capStr, metal):
-		return []string{"metal", "cpu"}
+		return []string{backendTokenMetal, "cpu"}
 	case strings.HasPrefix(capStr, darwinX86):
 		return []string{"darwin-x86", "cpu"}
 	case strings.HasPrefix(capStr, vulkan):
-		return []string{"vulkan", "cpu"}
+		return []string{vulkan, "cpu"}
 	default:
 		return []string{"cpu"}
 	}
 }
+
+// DetectedCapability returns the detected system capability string.
+// This can be used by the UI to display what capability was detected.
+func (s *SystemState) DetectedCapability() string {
+	return s.getSystemCapabilities()
+}
+
+// IsBackendCompatible checks if a backend (identified by name and URI) is compatible
+// with the current system capability. This function uses getSystemCapabilities to ensure
+// consistency with capability detection (including VRAM checks, environment overrides, etc.).
+func (s *SystemState) IsBackendCompatible(name, uri string) bool {
+	combined := strings.ToLower(name + " " + uri)
+	capability := s.getSystemCapabilities()
+
+	// Check for darwin/macOS-specific backends (mlx, metal, darwin)
+	isDarwinBackend := strings.Contains(combined, backendTokenDarwin) ||
+		strings.Contains(combined, backendTokenMLX) ||
+		strings.Contains(combined, backendTokenMetal)
+	if isDarwinBackend {
+		// Darwin backends require the system to be running on darwin with metal or darwin-x86 capability
+		return capability == metal || capability == darwinX86
+	}
+
+	// Check for NVIDIA L4T-specific backends (arm64 Linux with NVIDIA GPU)
+	// This must be checked before the general NVIDIA check as L4T backends
+	// may also contain "cuda" or "nvidia" in their names
+	isL4TBackend := strings.Contains(combined, backendTokenL4T)
+	if isL4TBackend {
+		return strings.HasPrefix(capability, nvidiaL4T)
+	}
+
+	// Check for NVIDIA/CUDA-specific backends (non-L4T)
+	isNvidiaBackend := strings.Contains(combined, backendTokenCUDA) ||
+		strings.Contains(combined, Nvidia)
+	if isNvidiaBackend {
+		// NVIDIA backends are compatible with nvidia, nvidia-cuda-12, nvidia-cuda-13, and l4t capabilities
+		return strings.HasPrefix(capability, Nvidia)
+	}
+
+	// Check for AMD/ROCm-specific backends
+	isAMDBackend := strings.Contains(combined, backendTokenROCM) ||
+		strings.Contains(combined, backendTokenHIP) ||
+		strings.Contains(combined, AMD)
+	if isAMDBackend {
+		return capability == AMD
+	}
+
+	// Check for Intel/SYCL-specific backends
+	isIntelBackend := strings.Contains(combined, backendTokenSYCL) ||
+		strings.Contains(combined, Intel)
+	if isIntelBackend {
+		return capability == Intel
+	}
+
+	// CPU backends are always compatible
+	return true
+}

swagger/docs.go

@@ -1198,6 +1198,30 @@ const docTemplate = `{
                 }
             }
         },
+        "/v1/messages": {
+            "post": {
+                "summary": "Generate a message response for the given messages and model.",
+                "parameters": [
+                    {
+                        "description": "query params",
+                        "name": "request",
+                        "in": "body",
+                        "required": true,
+                        "schema": {
+                            "$ref": "#/definitions/schema.AnthropicRequest"
+                        }
+                    }
+                ],
+                "responses": {
+                    "200": {
+                        "description": "Response",
+                        "schema": {
+                            "$ref": "#/definitions/schema.AnthropicResponse"
+                        }
+                    }
+                }
+            }
+        },
         "/v1/models": {
             "get": {
                 "summary": "List and describe the various models available in the API.",
@@ -1739,6 +1763,169 @@ const docTemplate = `{
                 }
             }
         },
+        "schema.AnthropicContentBlock": {
+            "type": "object",
+            "properties": {
+                "content": {},
+                "id": {
+                    "type": "string"
+                },
+                "input": {
+                    "type": "object",
+                    "additionalProperties": true
+                },
+                "is_error": {
+                    "type": "boolean"
+                },
+                "name": {
+                    "type": "string"
+                },
+                "source": {
+                    "$ref": "#/definitions/schema.AnthropicImageSource"
+                },
+                "text": {
+                    "type": "string"
+                },
+                "tool_use_id": {
+                    "type": "string"
+                },
+                "type": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicImageSource": {
+            "type": "object",
+            "properties": {
+                "data": {
+                    "type": "string"
+                },
+                "media_type": {
+                    "type": "string"
+                },
+                "type": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicMessage": {
+            "type": "object",
+            "properties": {
+                "content": {},
+                "role": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicRequest": {
+            "type": "object",
+            "properties": {
+                "max_tokens": {
+                    "type": "integer"
+                },
+                "messages": {
+                    "type": "array",
+                    "items": {
+                        "$ref": "#/definitions/schema.AnthropicMessage"
+                    }
+                },
+                "metadata": {
+                    "type": "object",
+                    "additionalProperties": {
+                        "type": "string"
+                    }
+                },
+                "model": {
+                    "type": "string"
+                },
+                "stop_sequences": {
+                    "type": "array",
+                    "items": {
+                        "type": "string"
+                    }
+                },
+                "stream": {
+                    "type": "boolean"
+                },
+                "system": {
+                    "type": "string"
+                },
+                "temperature": {
+                    "type": "number"
+                },
+                "tool_choice": {},
+                "tools": {
+                    "type": "array",
+                    "items": {
+                        "$ref": "#/definitions/schema.AnthropicTool"
+                    }
+                },
+                "top_k": {
+                    "type": "integer"
+                },
+                "top_p": {
+                    "type": "number"
+                }
+            }
+        },
+        "schema.AnthropicResponse": {
+            "type": "object",
+            "properties": {
+                "content": {
+                    "type": "array",
+                    "items": {
+                        "$ref": "#/definitions/schema.AnthropicContentBlock"
+                    }
+                },
+                "id": {
+                    "type": "string"
+                },
+                "model": {
+                    "type": "string"
+                },
+                "role": {
+                    "type": "string"
+                },
+                "stop_reason": {
+                    "type": "string"
+                },
+                "stop_sequence": {
+                    "type": "string"
+                },
+                "type": {
+                    "type": "string"
+                },
+                "usage": {
+                    "$ref": "#/definitions/schema.AnthropicUsage"
+                }
+            }
+        },
+        "schema.AnthropicTool": {
+            "type": "object",
+            "properties": {
+                "description": {
+                    "type": "string"
+                },
+                "input_schema": {
+                    "type": "object",
+                    "additionalProperties": true
+                },
+                "name": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicUsage": {
+            "type": "object",
+            "properties": {
+                "input_tokens": {
+                    "type": "integer"
+                },
+                "output_tokens": {
+                    "type": "integer"
+                }
+            }
+        },
         "schema.BackendMonitorRequest": {
             "type": "object",
             "properties": {
@@ -2229,6 +2416,10 @@ const docTemplate = `{
                     "description": "The message name (used for tools calls)",
                     "type": "string"
                 },
+                "reasoning": {
+                    "description": "Reasoning content extracted from \u003cthinking\u003e...\u003c/thinking\u003e tags",
+                    "type": "string"
+                },
                 "role": {
                     "description": "The message role",
                     "type": "string"

swagger/swagger.json

@@ -1191,6 +1191,30 @@
                 }
             }
         },
+        "/v1/messages": {
+            "post": {
+                "summary": "Generate a message response for the given messages and model.",
+                "parameters": [
+                    {
+                        "description": "query params",
+                        "name": "request",
+                        "in": "body",
+                        "required": true,
+                        "schema": {
+                            "$ref": "#/definitions/schema.AnthropicRequest"
+                        }
+                    }
+                ],
+                "responses": {
+                    "200": {
+                        "description": "Response",
+                        "schema": {
+                            "$ref": "#/definitions/schema.AnthropicResponse"
+                        }
+                    }
+                }
+            }
+        },
         "/v1/models": {
             "get": {
                 "summary": "List and describe the various models available in the API.",
@@ -1732,6 +1756,169 @@
                 }
             }
         },
+        "schema.AnthropicContentBlock": {
+            "type": "object",
+            "properties": {
+                "content": {},
+                "id": {
+                    "type": "string"
+                },
+                "input": {
+                    "type": "object",
+                    "additionalProperties": true
+                },
+                "is_error": {
+                    "type": "boolean"
+                },
+                "name": {
+                    "type": "string"
+                },
+                "source": {
+                    "$ref": "#/definitions/schema.AnthropicImageSource"
+                },
+                "text": {
+                    "type": "string"
+                },
+                "tool_use_id": {
+                    "type": "string"
+                },
+                "type": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicImageSource": {
+            "type": "object",
+            "properties": {
+                "data": {
+                    "type": "string"
+                },
+                "media_type": {
+                    "type": "string"
+                },
+                "type": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicMessage": {
+            "type": "object",
+            "properties": {
+                "content": {},
+                "role": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicRequest": {
+            "type": "object",
+            "properties": {
+                "max_tokens": {
+                    "type": "integer"
+                },
+                "messages": {
+                    "type": "array",
+                    "items": {
+                        "$ref": "#/definitions/schema.AnthropicMessage"
+                    }
+                },
+                "metadata": {
+                    "type": "object",
+                    "additionalProperties": {
+                        "type": "string"
+                    }
+                },
+                "model": {
+                    "type": "string"
+                },
+                "stop_sequences": {
+                    "type": "array",
+                    "items": {
+                        "type": "string"
+                    }
+                },
+                "stream": {
+                    "type": "boolean"
+                },
+                "system": {
+                    "type": "string"
+                },
+                "temperature": {
+                    "type": "number"
+                },
+                "tool_choice": {},
+                "tools": {
+                    "type": "array",
+                    "items": {
+                        "$ref": "#/definitions/schema.AnthropicTool"
+                    }
+                },
+                "top_k": {
+                    "type": "integer"
+                },
+                "top_p": {
+                    "type": "number"
+                }
+            }
+        },
+        "schema.AnthropicResponse": {
+            "type": "object",
+            "properties": {
+                "content": {
+                    "type": "array",
+                    "items": {
+                        "$ref": "#/definitions/schema.AnthropicContentBlock"
+                    }
+                },
+                "id": {
+                    "type": "string"
+                },
+                "model": {
+                    "type": "string"
+                },
+                "role": {
+                    "type": "string"
+                },
+                "stop_reason": {
+                    "type": "string"
+                },
+                "stop_sequence": {
+                    "type": "string"
+                },
+                "type": {
+                    "type": "string"
+                },
+                "usage": {
+                    "$ref": "#/definitions/schema.AnthropicUsage"
+                }
+            }
+        },
+        "schema.AnthropicTool": {
+            "type": "object",
+            "properties": {
+                "description": {
+                    "type": "string"
+                },
+                "input_schema": {
+                    "type": "object",
+                    "additionalProperties": true
+                },
+                "name": {
+                    "type": "string"
+                }
+            }
+        },
+        "schema.AnthropicUsage": {
+            "type": "object",
+            "properties": {
+                "input_tokens": {
+                    "type": "integer"
+                },
+                "output_tokens": {
+                    "type": "integer"
+                }
+            }
+        },
         "schema.BackendMonitorRequest": {
             "type": "object",
             "properties": {
@@ -2222,6 +2409,10 @@
                     "description": "The message name (used for tools calls)",
                     "type": "string"
                 },
+                "reasoning": {
+                    "description": "Reasoning content extracted from \u003cthinking\u003e...\u003c/thinking\u003e tags",
+                    "type": "string"
+                },
                 "role": {
                     "description": "The message role",
                     "type": "string"

swagger/swagger.yaml

@@ -239,6 +239,114 @@ definitions:
       start:
         type: number
     type: object
+  schema.AnthropicContentBlock:
+    properties:
+      content: {}
+      id:
+        type: string
+      input:
+        additionalProperties: true
+        type: object
+      is_error:
+        type: boolean
+      name:
+        type: string
+      source:
+        $ref: '#/definitions/schema.AnthropicImageSource'
+      text:
+        type: string
+      tool_use_id:
+        type: string
+      type:
+        type: string
+    type: object
+  schema.AnthropicImageSource:
+    properties:
+      data:
+        type: string
+      media_type:
+        type: string
+      type:
+        type: string
+    type: object
+  schema.AnthropicMessage:
+    properties:
+      content: {}
+      role:
+        type: string
+    type: object
+  schema.AnthropicRequest:
+    properties:
+      max_tokens:
+        type: integer
+      messages:
+        items:
+          $ref: '#/definitions/schema.AnthropicMessage'
+        type: array
+      metadata:
+        additionalProperties:
+          type: string
+        type: object
+      model:
+        type: string
+      stop_sequences:
+        items:
+          type: string
+        type: array
+      stream:
+        type: boolean
+      system:
+        type: string
+      temperature:
+        type: number
+      tool_choice: {}
+      tools:
+        items:
+          $ref: '#/definitions/schema.AnthropicTool'
+        type: array
+      top_k:
+        type: integer
+      top_p:
+        type: number
+    type: object
+  schema.AnthropicResponse:
+    properties:
+      content:
+        items:
+          $ref: '#/definitions/schema.AnthropicContentBlock'
+        type: array
+      id:
+        type: string
+      model:
+        type: string
+      role:
+        type: string
+      stop_reason:
+        type: string
+      stop_sequence:
+        type: string
+      type:
+        type: string
+      usage:
+        $ref: '#/definitions/schema.AnthropicUsage'
+    type: object
+  schema.AnthropicTool:
+    properties:
+      description:
+        type: string
+      input_schema:
+        additionalProperties: true
+        type: object
+      name:
+        type: string
+    type: object
+  schema.AnthropicUsage:
+    properties:
+      input_tokens:
+        type: integer
+      output_tokens:
+        type: integer
+    type: object
   schema.BackendMonitorRequest:
     properties:
       model:
@@ -573,6 +681,9 @@ definitions:
       name:
         description: The message name (used for tools calls)
         type: string
+      reasoning:
+        description: Reasoning content extracted from <thinking>...</thinking> tags
+        type: string
       role:
         description: The message role
         type: string
@@ -1813,6 +1924,21 @@ paths:
           schema:
             $ref: '#/definitions/schema.OpenAIResponse'
       summary: Stream MCP chat completions with reasoning, tool calls, and results
+  /v1/messages:
+    post:
+      parameters:
+      - description: query params
+        in: body
+        name: request
+        required: true
+        schema:
+          $ref: '#/definitions/schema.AnthropicRequest'
+      responses:
+        "200":
+          description: Response
+          schema:
+            $ref: '#/definitions/schema.AnthropicResponse'
+      summary: Generate a message response for the given messages and model.
   /v1/models:
     get:
       responses:

tests/e2e/e2e_anthropic_test.go

@@ -0,0 +1,375 @@
+package e2e_test
+
+import (
+	"context"
+
+	"github.com/anthropics/anthropic-sdk-go"
+	"github.com/anthropics/anthropic-sdk-go/option"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("Anthropic API E2E test", func() {
+	var client anthropic.Client
+
+	Context("API with Anthropic SDK", func() {
+		BeforeEach(func() {
+			// Create Anthropic client pointing to LocalAI
+			client = anthropic.NewClient(
+				option.WithBaseURL(localAIURL),
+				option.WithAPIKey("test-api-key"), // LocalAI doesn't require a real API key
+			)
+
+			// Wait for API to be ready by attempting a simple request
+			Eventually(func() error {
+				_, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 10,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("Hi")),
+					},
+				})
+				return err
+			}, "2m").ShouldNot(HaveOccurred())
+		})
+
+		Context("Non-streaming responses", func() {
+			It("generates a response for a simple message", func() {
+				message, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("How much is 2+2? Reply with just the number.")),
+					},
+				})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(message.Content).ToNot(BeEmpty())
+				// Role is a constant type that defaults to "assistant"
+				Expect(string(message.Role)).To(Equal("assistant"))
+				Expect(message.StopReason).To(Equal(anthropic.MessageStopReasonEndTurn))
+				Expect(string(message.Type)).To(Equal("message"))
+
+				// Check that content contains text block with expected answer
+				Expect(len(message.Content)).To(BeNumerically(">=", 1))
+				textBlock := message.Content[0]
+				Expect(string(textBlock.Type)).To(Equal("text"))
+				Expect(textBlock.Text).To(Or(ContainSubstring("4"), ContainSubstring("four")))
+			})
+
+			It("handles system prompts", func() {
+				message, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					System: []anthropic.TextBlockParam{
+						{Text: "You are a helpful assistant. Always respond in uppercase letters."},
+					},
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("Say hello")),
+					},
+				})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(message.Content).ToNot(BeEmpty())
+				Expect(len(message.Content)).To(BeNumerically(">=", 1))
+			})
+
+			It("returns usage information", func() {
+				message, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 100,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("Hello")),
+					},
+				})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(message.Usage.InputTokens).To(BeNumerically(">", 0))
+				Expect(message.Usage.OutputTokens).To(BeNumerically(">", 0))
+			})
+		})
+
+		Context("Streaming responses", func() {
+			It("streams tokens for a simple message", func() {
+				stream := client.Messages.NewStreaming(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("Count from 1 to 5")),
+					},
+				})
+
+				message := anthropic.Message{}
+				eventCount := 0
+				hasContentDelta := false
+
+				for stream.Next() {
+					event := stream.Current()
+					err := message.Accumulate(event)
+					Expect(err).ToNot(HaveOccurred())
+					eventCount++
+
+					// Check for content block delta events
+					switch event.AsAny().(type) {
+					case anthropic.ContentBlockDeltaEvent:
+						hasContentDelta = true
+					}
+				}
+
+				Expect(stream.Err()).ToNot(HaveOccurred())
+				Expect(eventCount).To(BeNumerically(">", 0))
+				Expect(hasContentDelta).To(BeTrue())
+
+				// Check accumulated message
+				Expect(message.Content).ToNot(BeEmpty())
+				// Role is a constant type that defaults to "assistant"
+				Expect(string(message.Role)).To(Equal("assistant"))
+			})
+
+			It("streams with system prompt", func() {
+				stream := client.Messages.NewStreaming(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					System: []anthropic.TextBlockParam{
+						{Text: "You are a helpful assistant."},
+					},
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("Say hello")),
+					},
+				})
+
+				message := anthropic.Message{}
+				for stream.Next() {
+					event := stream.Current()
+					err := message.Accumulate(event)
+					Expect(err).ToNot(HaveOccurred())
+				}
+
+				Expect(stream.Err()).ToNot(HaveOccurred())
+				Expect(message.Content).ToNot(BeEmpty())
+			})
+		})
+
+		Context("Tool calling", func() {
+			It("handles tool calls in non-streaming mode", func() {
+				message, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("What's the weather like in San Francisco?")),
+					},
+					Tools: []anthropic.ToolParam{
+						{
+							Name:        "get_weather",
+							Description: anthropic.F("Get the current weather in a given location"),
+							InputSchema: anthropic.F(map[string]interface{}{
+								"type": "object",
+								"properties": map[string]interface{}{
+									"location": map[string]interface{}{
+										"type":        "string",
+										"description": "The city and state, e.g. San Francisco, CA",
+									},
+								},
+								"required": []string{"location"},
+							}),
+						},
+					},
+				})
+
+				Expect(err).ToNot(HaveOccurred())
+				Expect(message.Content).ToNot(BeEmpty())
+
+				// The model must use tools - find the tool use in the response
+				hasToolUse := false
+				for _, block := range message.Content {
+					if block.Type == anthropic.ContentBlockTypeToolUse {
+						hasToolUse = true
+						Expect(block.Name).To(Equal("get_weather"))
+						Expect(block.ID).ToNot(BeEmpty())
+						// Verify that input contains location
+						inputMap, ok := block.Input.(map[string]interface{})
+						Expect(ok).To(BeTrue())
+						_, hasLocation := inputMap["location"]
+						Expect(hasLocation).To(BeTrue())
+					}
+				}
+
+				// Model must have called the tool
+				Expect(hasToolUse).To(BeTrue(), "Model should have called the get_weather tool")
+				Expect(message.StopReason).To(Equal(anthropic.MessageStopReasonToolUse))
+			})
+
+			It("handles tool_choice parameter", func() {
+				message, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("Tell me about the weather")),
+					},
+					Tools: []anthropic.ToolParam{
+						{
+							Name:        "get_weather",
+							Description: anthropic.F("Get the current weather"),
+							InputSchema: anthropic.F(map[string]interface{}{
+								"type": "object",
+								"properties": map[string]interface{}{
+									"location": map[string]interface{}{
+										"type": "string",
+									},
+								},
+							}),
+						},
+					},
+					ToolChoice: anthropic.F[anthropic.ToolChoiceUnionParam](
+						anthropic.ToolChoiceAutoParam{
+							Type: anthropic.F(anthropic.ToolChoiceAutoTypeAuto),
+						},
+					),
+				})
+
+				Expect(err).ToNot(HaveOccurred())
+				Expect(message.Content).ToNot(BeEmpty())
+			})
+
+			It("handles tool results in messages", func() {
+				// First, make a request that should trigger a tool call
+				firstMessage, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("What's the weather in SF?")),
+					},
+					Tools: []anthropic.ToolParam{
+						{
+							Name:        "get_weather",
+							Description: anthropic.F("Get weather"),
+							InputSchema: anthropic.F(map[string]interface{}{
+								"type": "object",
+								"properties": map[string]interface{}{
+									"location": map[string]interface{}{"type": "string"},
+								},
+							}),
+						},
+					},
+				})
+
+				Expect(err).ToNot(HaveOccurred())
+
+				// Find the tool use block - model must call the tool
+				var toolUseID string
+				var toolUseName string
+				for _, block := range firstMessage.Content {
+					if block.Type == anthropic.ContentBlockTypeToolUse {
+						toolUseID = block.ID
+						toolUseName = block.Name
+						break
+					}
+				}
+
+				// Model must have called the tool
+				Expect(toolUseID).ToNot(BeEmpty(), "Model should have called the get_weather tool")
+
+				// Send back a tool result and verify it's handled correctly
+				secondMessage, err := client.Messages.New(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("What's the weather in SF?")),
+						anthropic.NewAssistantMessage(firstMessage.Content...),
+						anthropic.NewUserMessage(
+							anthropic.NewToolResultBlock(toolUseID, "Sunny, 72°F", false),
+						),
+					},
+					Tools: []anthropic.ToolParam{
+						{
+							Name:        toolUseName,
+							Description: anthropic.F("Get weather"),
+							InputSchema: anthropic.F(map[string]interface{}{
+								"type": "object",
+								"properties": map[string]interface{}{
+									"location": map[string]interface{}{"type": "string"},
+								},
+							}),
+						},
+					},
+				})
+
+				Expect(err).ToNot(HaveOccurred())
+				Expect(secondMessage.Content).ToNot(BeEmpty())
+			})
+
+			It("handles tool calls in streaming mode", func() {
+				stream := client.Messages.NewStreaming(context.TODO(), anthropic.MessageNewParams{
+					Model:     "gpt-4",
+					MaxTokens: 1024,
+					Messages: []anthropic.MessageParam{
+						anthropic.NewUserMessage(anthropic.NewTextBlock("What's the weather like in San Francisco?")),
+					},
+					Tools: []anthropic.ToolParam{
+						{
+							Name:        "get_weather",
+							Description: anthropic.F("Get the current weather in a given location"),
+							InputSchema: anthropic.F(map[string]interface{}{
+								"type": "object",
+								"properties": map[string]interface{}{
+									"location": map[string]interface{}{
+										"type":        "string",
+										"description": "The city and state, e.g. San Francisco, CA",
+									},
+								},
+								"required": []string{"location"},
+							}),
+						},
+					},
+				})
+
+				message := anthropic.Message{}
+				eventCount := 0
+				hasToolUseBlock := false
+				hasContentBlockStart := false
+				hasContentBlockDelta := false
+				hasContentBlockStop := false
+
+				for stream.Next() {
+					event := stream.Current()
+					err := message.Accumulate(event)
+					Expect(err).ToNot(HaveOccurred())
+					eventCount++
+
+					// Check for different event types related to tool use
+					switch e := event.AsAny().(type) {
+					case anthropic.ContentBlockStartEvent:
+						hasContentBlockStart = true
+						if e.ContentBlock.Type == anthropic.ContentBlockTypeToolUse {
+							hasToolUseBlock = true
+						}
+					case anthropic.ContentBlockDeltaEvent:
+						hasContentBlockDelta = true
+					case anthropic.ContentBlockStopEvent:
+						hasContentBlockStop = true
+					}
+				}
+
+				Expect(stream.Err()).ToNot(HaveOccurred())
+				Expect(eventCount).To(BeNumerically(">", 0))
+
+				// Verify streaming events were emitted
+				Expect(hasContentBlockStart).To(BeTrue(), "Should have content_block_start event")
+				Expect(hasContentBlockDelta).To(BeTrue(), "Should have content_block_delta event")
+				Expect(hasContentBlockStop).To(BeTrue(), "Should have content_block_stop event")
+
+				// Check accumulated message has tool use
+				Expect(message.Content).ToNot(BeEmpty())
+				
+				// Model must have called the tool
+				foundToolUse := false
+				for _, block := range message.Content {
+					if block.Type == anthropic.ContentBlockTypeToolUse {
+						foundToolUse = true
+						Expect(block.Name).To(Equal("get_weather"))
+						Expect(block.ID).ToNot(BeEmpty())
+					}
+				}
+				Expect(foundToolUse).To(BeTrue(), "Model should have called the get_weather tool in streaming mode")
+				Expect(message.StopReason).To(Equal(anthropic.MessageStopReasonToolUse))
+			})
+		})
+	})
+})