feat(gallery): add Qwen3.8 iMatrix NVFP4 variant

Fix the existing NVFP4 variant to use its downloaded BF16 projector.

Assisted-by: Codex:GPT-6 [shell]

docs/content/features/model-gallery.md

@@ -1112,3 +1112,20 @@ is still running is reported as queued until the installer picks it up:
 A job ID is queryable from the moment `/models/apply` returns it, so a `404`/`500`
 from this endpoint means the ID is genuinely unknown rather than merely waiting
 its turn.
+
+## Qwen3.8-27B iMatrix NVFP4
+
+The `qwen3.8-27b-imatrix-nvfp4-mtp` variant uses cdiamond's mixed-precision
+NVFP4 GGUF with llama.cpp. It includes the F16 vision projector and enables
+MTP with one draft token per step. The default context is 32,768 tokens.
+
+Select this build explicitly with:
+
+```sh
+local-ai models install qwen3.8-27b-q4 --variant qwen3.8-27b-imatrix-nvfp4-mtp
+```
+
+The model and projector total about 18.1 GB; inference requires additional
+memory for context and working buffers. See the
+[publisher's model card](https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF)
+for the quantization recipe and hardware-specific measurements.

gallery/index.yaml

@@ -3,6 +3,68 @@
 # written by scripts/models/cmd/modelcatalog and stays only so releases before
 # the cutover keep working. CI fails when it differs from the generated copy:
 #   go run ./scripts/models/cmd/modelcatalog --library gallery/models/index.yaml --out gallery/index.yaml
+- url: github:mudler/LocalAI/gallery/virtual.yaml@master
+  name: qwen3.8-27b-imatrix-nvfp4-mtp
+  description: |
+    Qwen3.8-27B in cdiamond's iMatrix-calibrated mixed-precision NVFP4
+    build. Selected attention and feed-forward tensors retain higher
+    precision. Includes the F16 vision projector and enables the embedded
+    MTP draft head with one draft token per step.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/Qwen/Qwen3.8-27B
+    - https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF
+  icon: https://qianwen-res.oss-cn-beijing.aliyuncs.com/logo_qwen.jpg
+  tags:
+    - llm
+    - gguf
+    - gpu
+    - qwen
+    - reasoning
+    - thinking
+    - coding
+    - agent
+    - tools
+    - vision
+    - multimodal
+    - long-context
+    - nvfp4
+    - mtp
+    - speculative
+  files:
+    - filename: llama-cpp/models/qwen3.8-27b/Qwen3.8-27B-iMatrix-NVFP4-MTP.gguf
+      sha256: 15b6bb8d0eb15cb6b081c562f8daea86330ec64174cb0f9e1655f3d467484007
+      uri: https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF/resolve/ac343e8f44caef0896f79d372ecc07ef7ab34ec8/Qwen3.8-27B-iMatrix-NVFP4-MTP.gguf
+    - filename: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-iMatrix-F16.gguf
+      sha256: cbb841a9ee0636b2ec172f5bb8df2ea8dfeb01e90fe7c6126581d662a0b4e43e
+      uri: https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF/resolve/ac343e8f44caef0896f79d372ecc07ef7ab34ec8/mmproj-Qwen3.8-27B-F16.gguf
+  overrides:
+    backend: llama-cpp
+    context_size: 32768
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+      - vision
+    mmproj: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-iMatrix-F16.gguf
+    options:
+      - use_jinja:true
+      - spec_type:draft-mtp
+      - spec_n_max:1
+      - spec_p_min:0.75
+    parameters:
+      min_p: 0
+      model: llama-cpp/models/qwen3.8-27b/Qwen3.8-27B-iMatrix-NVFP4-MTP.gguf
+      presence_penalty: 0
+      repeat_penalty: 1
+      temperature: 1
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  variants: []
 - url: github:mudler/LocalAI/gallery/virtual.yaml@master
   name: adavar-3b-q4
   description: |
@@ -7088,6 +7150,7 @@
   variants:
     - model: qwen3.8-27b-q4-mtp
     - model: qwen3.8-27b-nvfp4-mtp
+    - model: qwen3.8-27b-imatrix-nvfp4-mtp
     - model: qwen3.8-27b-ridge
     - model: qwen3.8-27b-gsq-rco-iq2-xs
     - model: qwen3.8-27b-gsq-rco-iq2-s
@@ -7204,7 +7267,7 @@
     known_usecases:
       - chat
       - vision
-    mmproj: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-Q8_0.gguf
+    mmproj: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-BF16.gguf
     options:
       - use_jinja:true
       - spec_type:draft-mtp

gallery/models/index.yaml

@@ -4,6 +4,76 @@
 # changes and never edit it by hand. CI fails when it is stale. A build from
 # master reads it as the default model source; a tagged build reads the gallery
 # until the release cutover.
+- kind: model
+  id: qwen3.8-27b-imatrix-nvfp4-mtp
+  name: qwen3.8-27b-imatrix-nvfp4-mtp
+  version: 0.0.0-gallery.6dcacd1339cc
+  description: |
+    Qwen3.8-27B in cdiamond's iMatrix-calibrated mixed-precision NVFP4
+    build. Selected attention and feed-forward tensors retain higher
+    precision. Includes the F16 vision projector and enables the embedded
+    MTP draft head with one draft token per step.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/Qwen/Qwen3.8-27B
+    - https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF
+  tags:
+    - llm
+    - gguf
+    - gpu
+    - qwen
+    - reasoning
+    - thinking
+    - coding
+    - agent
+    - tools
+    - vision
+    - multimodal
+    - long-context
+    - nvfp4
+    - mtp
+    - speculative
+  icon: https://qianwen-res.oss-cn-beijing.aliyuncs.com/logo_qwen.jpg
+  channel: gallery
+  policy:
+    requireSignature: false
+  schemaVersion: 1
+  model:
+    url: github:mudler/LocalAI/gallery/virtual.yaml@master
+    files:
+      - filename: llama-cpp/models/qwen3.8-27b/Qwen3.8-27B-iMatrix-NVFP4-MTP.gguf
+        sha256: 15b6bb8d0eb15cb6b081c562f8daea86330ec64174cb0f9e1655f3d467484007
+        uri: https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF/resolve/ac343e8f44caef0896f79d372ecc07ef7ab34ec8/Qwen3.8-27B-iMatrix-NVFP4-MTP.gguf
+      - filename: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-iMatrix-F16.gguf
+        sha256: cbb841a9ee0636b2ec172f5bb8df2ea8dfeb01e90fe7c6126581d662a0b4e43e
+        uri: https://huggingface.co/cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF/resolve/ac343e8f44caef0896f79d372ecc07ef7ab34ec8/mmproj-Qwen3.8-27B-F16.gguf
+    overrides:
+      backend: llama-cpp
+      context_size: 32768
+      function:
+        automatic_tool_parsing_fallback: true
+        grammar:
+          disable: true
+      known_usecases:
+        - chat
+        - vision
+      mmproj: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-iMatrix-F16.gguf
+      options:
+        - use_jinja:true
+        - spec_type:draft-mtp
+        - spec_n_max:1
+        - spec_p_min:0.75
+      parameters:
+        min_p: 0
+        model: llama-cpp/models/qwen3.8-27b/Qwen3.8-27B-iMatrix-NVFP4-MTP.gguf
+        presence_penalty: 0
+        repeat_penalty: 1
+        temperature: 1
+        top_k: 20
+        top_p: 0.95
+      template:
+        use_tokenizer_template: true
+    variants: []
 - kind: model
   id: adavar-3b-q4
   name: adavar-3b-q4
@@ -8209,6 +8279,7 @@
     variants:
       - model: qwen3.8-27b-q4-mtp
       - model: qwen3.8-27b-nvfp4-mtp
+      - model: qwen3.8-27b-imatrix-nvfp4-mtp
       - model: qwen3.8-27b-ridge
       - model: qwen3.8-27b-gsq-rco-iq2-xs
       - model: qwen3.8-27b-gsq-rco-iq2-s
@@ -8341,7 +8412,7 @@
       known_usecases:
         - chat
         - vision
-      mmproj: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-Q8_0.gguf
+      mmproj: llama-cpp/mmproj/qwen3.8-27b/mmproj-Qwen3.8-27B-BF16.gguf
       options:
         - use_jinja:true
         - spec_type:draft-mtp