feat(gallery): add Qwen3.8 GSQ-RCO variants (#11787)

Add three llama.cpp-compatible mixed quantizations from ISTA DASLab. These builds give Qwen3.8-27B users an 8.4 to 10.1 GB weight tier with the shared vision projector. Assisted-by: Codex:gpt-5 Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
2026-09-12 22:33:54 -04:00 · 2026-08-30 20:10:22 +02:00
1 parent 9c7c9974eb
commit dd4e75983d
@@ -1630,6 +1630,9 @@
  variants:
    - model: qwen3.8-27b-q4-mtp
    - model: qwen3.8-27b-ridge
+    - model: qwen3.8-27b-gsq-rco-iq2-xs
+    - model: qwen3.8-27b-gsq-rco-iq2-s
+    - model: qwen3.8-27b-gsq-rco-iq3-xxs
    - model: qwen3.8-27b-q8
  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
  urls:
@@ -1845,6 +1848,129 @@
    - filename: llama-cpp/mmproj/qwen3.8-27b-ridge/mmproj-Qwen3.8-27B-BF16.gguf
      uri: huggingface://empero-ai/Qwen3.8-27B-Ridge-GGUF/mmproj-Qwen3.8-27B-BF16.gguf
      sha256: 52228402ce4823f10705d901813cd43ced71859524cf2d8bf83305ad6b7dcbc2
+- !!merge <<: *qwen3-8-27b
+  name: "qwen3.8-27b-gsq-rco-iq2-xs"
+  variants: []
+  urls:
+    - https://huggingface.co/Qwen/Qwen3.8-27B
+    - https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF
+  description: |
+    Qwen3.8-27B with ISTA DASLab's smallest GSQ-RCO mixed quantization. The
+    2.50-bit IQ2_XS model uses 8.4 GB for the weights and retains the shared
+    BF16 vision projector.
+  last_checked: "2026-08-30"
+  overrides:
+    backend: llama-cpp
+    context_size: 262144
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+      - vision
+    mmproj: llama-cpp/mmproj/qwen3.8-27b-gsq-rco/mmproj-Qwen3.8-27B-BF16.gguf
+    options:
+      - use_jinja:true
+    parameters:
+      min_p: 0
+      model: llama-cpp/models/qwen3.8-27b-gsq-rco/Qwen3.8-27B-GSQ-RCO-IQ2_XS.gguf
+      presence_penalty: 0
+      repeat_penalty: 1
+      temperature: 1
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/qwen3.8-27b-gsq-rco/Qwen3.8-27B-GSQ-RCO-IQ2_XS.gguf
+      uri: huggingface://ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/Qwen3.8-27B-GSQ-RCO-IQ2_XS.gguf
+      sha256: f0ae5006da0ce6225935339e4e989369f94de95d2263cf969519f8420c9ae02c
+    - filename: llama-cpp/mmproj/qwen3.8-27b-gsq-rco/mmproj-Qwen3.8-27B-BF16.gguf
+      uri: huggingface://ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/mmproj-Qwen3.8-27B-BF16.gguf
+      sha256: 13cb7bebccbd04afc8f4090cb949ecf8937cdf7377c5799b1a0c594e7c0d3e16
+- !!merge <<: *qwen3-8-27b
+  name: "qwen3.8-27b-gsq-rco-iq2-s"
+  variants: []
+  urls:
+    - https://huggingface.co/Qwen/Qwen3.8-27B
+    - https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF
+  description: |
+    Qwen3.8-27B with ISTA DASLab's 2.75-bit GSQ-RCO IQ2_S mixed quantization.
+    The 9.3 GB model targets higher fidelity than IQ2_XS and retains the shared
+    BF16 vision projector.
+  last_checked: "2026-08-30"
+  overrides:
+    backend: llama-cpp
+    context_size: 262144
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+      - vision
+    mmproj: llama-cpp/mmproj/qwen3.8-27b-gsq-rco/mmproj-Qwen3.8-27B-BF16.gguf
+    options:
+      - use_jinja:true
+    parameters:
+      min_p: 0
+      model: llama-cpp/models/qwen3.8-27b-gsq-rco/Qwen3.8-27B-GSQ-RCO-IQ2_S.gguf
+      presence_penalty: 0
+      repeat_penalty: 1
+      temperature: 1
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/qwen3.8-27b-gsq-rco/Qwen3.8-27B-GSQ-RCO-IQ2_S.gguf
+      uri: huggingface://ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/Qwen3.8-27B-GSQ-RCO-IQ2_S.gguf
+      sha256: 16c9802111aa9ef3acde465188d6d601f8db128ee3d828ad983a5caca4135ecb
+    - filename: llama-cpp/mmproj/qwen3.8-27b-gsq-rco/mmproj-Qwen3.8-27B-BF16.gguf
+      uri: huggingface://ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/mmproj-Qwen3.8-27B-BF16.gguf
+      sha256: 13cb7bebccbd04afc8f4090cb949ecf8937cdf7377c5799b1a0c594e7c0d3e16
+- !!merge <<: *qwen3-8-27b
+  name: "qwen3.8-27b-gsq-rco-iq3-xxs"
+  variants: []
+  urls:
+    - https://huggingface.co/Qwen/Qwen3.8-27B
+    - https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF
+  description: |
+    Qwen3.8-27B with ISTA DASLab's recommended 3.00-bit GSQ-RCO IQ3_XXS mixed
+    quantization. The 10.1 GB model provides the highest quality in this set
+    and retains the shared BF16 vision projector.
+  last_checked: "2026-08-30"
+  overrides:
+    backend: llama-cpp
+    context_size: 262144
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+      - vision
+    mmproj: llama-cpp/mmproj/qwen3.8-27b-gsq-rco/mmproj-Qwen3.8-27B-BF16.gguf
+    options:
+      - use_jinja:true
+    parameters:
+      min_p: 0
+      model: llama-cpp/models/qwen3.8-27b-gsq-rco/Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf
+      presence_penalty: 0
+      repeat_penalty: 1
+      temperature: 1
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/qwen3.8-27b-gsq-rco/Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf
+      uri: huggingface://ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf
+      sha256: fdfcb6a29b11188956dfbfd904223588a6c1b77eb250c3e8a36e1bd269df91f7
+    - filename: llama-cpp/mmproj/qwen3.8-27b-gsq-rco/mmproj-Qwen3.8-27B-BF16.gguf
+      uri: huggingface://ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/mmproj-Qwen3.8-27B-BF16.gguf
+      sha256: 13cb7bebccbd04afc8f4090cb949ecf8937cdf7377c5799b1a0c594e7c0d3e16
 - &qwen3-8-9b
  name: "qwen3.8-9b-q4"
  variants: