feat(gallery): tag vllm-cpp entries by capability and add decision and vision models

laya declares the systemone usecase instead of chat. The gated Qwen3.6 27B
NVFP4 entries gain vision; the 35B-A3B entries gain it as experimental
because image input is not token-gated. Adds GLiNER2.5-Decide and
Qwen3-VL-4B. A guard test keeps capability tags and known_usecases in
agreement for every vllm-cpp entry.

Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

core/gallery/vllm_cpp_tags_test.go

@@ -0,0 +1,48 @@
+package gallery_test
+
+import (
+	"fmt"
+	"slices"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+)
+
+// A gallery tag that names a capability is what users filter on, and
+// known_usecases is what the server routes on. When they disagree, the entry
+// is listed under a filter it cannot serve, or is hidden from one it can.
+var _ = Describe("gallery/index.yaml vllm-cpp capability tags", func() {
+	It("keeps capability tags and known_usecases in agreement", func() {
+		entries, err := loadGalleryIndex()
+		Expect(err).ToNot(HaveOccurred())
+
+		tagToFlag := map[string]config.ModelConfigUsecase{
+			"systemone":      config.FLAG_SYSTEMONE,
+			"vision":         config.FLAG_VISION,
+			"token-classify": config.FLAG_TOKEN_CLASSIFY,
+			"scoring":        config.FLAG_SCORE,
+		}
+
+		var violations []string
+		seen := 0
+		for i := range entries {
+			e := &entries[i]
+			if backend, _ := e.Overrides["backend"].(string); backend != "vllm-cpp" {
+				continue
+			}
+			seen++
+			declared := e.GetKnownUsecases()
+			for tag, flag := range tagToFlag {
+				tagged := slices.Contains(e.Tags, tag)
+				has := declared != nil && *declared&flag == flag
+				if tagged != has {
+					violations = append(violations, fmt.Sprintf("%s: tag %q present=%v but known_usecases declares it=%v", e.Name, tag, tagged, has))
+				}
+			}
+		}
+		Expect(seen).To(BeNumerically(">", 0))
+		Expect(violations).To(BeEmpty())
+	})
+})

gallery/index.yaml

@@ -19360,6 +19360,7 @@
     - qwen3.6
     - nvfp4
     - vllm-cpp
+    - vision
     - tool-calling
     - reasoning
     - gpu
@@ -19372,6 +19373,7 @@
     known_usecases:
       - chat
       - completion
+      - vision
     # Tool calls and the <think> split are parsed by the engine's own streaming
     # parsers, so LocalAI's Go-side grammar path stays out of the way.
     function:
@@ -19429,6 +19431,7 @@
     - qwen3.6
     - nvfp4
     - vllm-cpp
+    - vision
     - speculative-decoding
     - mtp
     - tool-calling
@@ -19442,6 +19445,7 @@
     known_usecases:
       - chat
       - completion
+      - vision
     function:
       grammar:
         disable: true
@@ -19497,6 +19501,7 @@
     - qwen3.6
     - nvfp4
     - vllm-cpp
+    - vision
     - speculative-decoding
     - dflash
     - tool-calling
@@ -19510,6 +19515,7 @@
     known_usecases:
       - chat
       - completion
+      - vision
     function:
       grammar:
         disable: true
@@ -19557,6 +19563,9 @@
     with roughly 3B parameters active per token, so it reads like a much larger
     model while costing about as much per token as a small one.
 
+    Image input is implemented in the engine but is not token-gated against
+    vLLM yet, so the vision usecase on this entry is experimental.
+
     This is the engine's gated MoE checkpoint: token-for-token identical to vLLM
     over the 315-prompt battery on both the synchronous and asynchronous paths,
     at 0.92x to 0.97x vLLM's throughput from concurrency 1 to 32.
@@ -19575,6 +19584,8 @@
     - moe
     - nvfp4
     - vllm-cpp
+    - vision
+    - experimental
     - tool-calling
     - reasoning
     - gpu
@@ -19587,6 +19598,7 @@
     known_usecases:
       - chat
       - completion
+      - vision
     function:
       grammar:
         disable: true
@@ -19615,6 +19627,9 @@
   description: |
     Qwen3.6-35B-A3B NVFP4 on vllm.cpp with MTP speculative decoding enabled.
 
+    Image input is implemented in the engine but is not token-gated against
+    vLLM yet, so the vision usecase on this entry is experimental.
+
     The draft head ships inside the checkpoint's own mtp.* tensors, so there is
     no second model to download. On this model the speculative path is
     token-exact against speculation-off on both the synchronous and asynchronous
@@ -19632,6 +19647,8 @@
     - moe
     - nvfp4
     - vllm-cpp
+    - vision
+    - experimental
     - speculative-decoding
     - mtp
     - tool-calling
@@ -19645,6 +19662,7 @@
     known_usecases:
       - chat
       - completion
+      - vision
     function:
       grammar:
         disable: true
@@ -63645,7 +63663,7 @@
   overrides:
     backend: vllm-cpp
     known_usecases:
-      - chat
+      - systemone
     parameters:
       model: convaiinnovations/laya
   artifacts:
@@ -63654,6 +63672,90 @@
       source:
         type: huggingface
         repo: convaiinnovations/laya
+- name: gliner25-decide-vllm-cpp
+  url: github:mudler/LocalAI/gallery/virtual.yaml@master
+  urls:
+    - https://huggingface.co/fastino/GLiNER2.5-Decide
+    - https://github.com/mudler/vllm.cpp
+  description: |
+    GLiNER2.5-Decide is a DeBERTa-v3-large encoder with a classification head
+    that answers typed decision questions over a state text in one forward
+    pass. It never generates text, so there is nothing to parse.
+
+    In LocalAI, serve it via POST /v1/systemone. The vllm.cpp engine runs the
+    decision pipeline (choice, noul and score question types) through the
+    vllm_decide C ABI. This is the decision model, not the zero-shot NER model:
+    use the gliner2.5 entry for entity extraction. F32 weights, about 2 GB.
+    The weights are pinned to a revision so the entry keeps serving the
+    checkpoint it was checked against.
+  license: apache-2.0
+  tags:
+    - decision
+    - systemone
+    - vllm-cpp
+    - cpu
+    - gpu
+  size: 2GB
+  last_checked: "2026-09-30"
+  overrides:
+    backend: vllm-cpp
+    known_usecases:
+      - systemone
+    parameters:
+      model: fastino/GLiNER2.5-Decide
+  artifacts:
+    - name: model
+      target: model
+      source:
+        type: huggingface
+        repo: fastino/GLiNER2.5-Decide
+        revision: 5a7adf72a23b4d311abae6ce050d7f0012bb3416
+- name: qwen3-vl-4b-vllm-cpp
+  url: github:mudler/LocalAI/gallery/virtual.yaml@master
+  urls:
+    - https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct
+    - https://github.com/mudler/vllm.cpp
+  description: |
+    Qwen3-VL-4B-Instruct on vllm.cpp, in bf16: a small vision-language model
+    that takes images alongside text. In the engine's correctness battery the
+    image path matches vLLM token for token, and video input is a near tie.
+
+    Roughly 9 GB of weights plus KV cache at the context configured here. It
+    runs where the flagship NVFP4 checkpoints cannot, including plain CPU.
+  license: apache-2.0
+  tags:
+    - llm
+    - vision
+    - multimodal
+    - qwen
+    - qwen3-vl
+    - vllm-cpp
+    - cpu
+    - gpu
+  size: 9GB
+  last_checked: "2026-09-30"
+  overrides:
+    backend: vllm-cpp
+    known_usecases:
+      - chat
+      - completion
+      - vision
+    template:
+      use_tokenizer_template: true
+    context_size: 8192
+    engine_args:
+      block_size: 32
+      num_blocks: 512
+      max_num_seqs: 4
+    parameters:
+      model: Qwen/Qwen3-VL-4B-Instruct
+  artifacts:
+    - name: model
+      target: model
+      source:
+        type: huggingface
+        repo: Qwen/Qwen3-VL-4B-Instruct
+        revision: ebb281ec70b05090aa6165b016eac8ec08e71b17
 - name: cua-s1-forms-vllm-cpp
   url: github:mudler/LocalAI/gallery/virtual.yaml@master
   urls: