From 70a2bd2c21cd18e39e8bdf3a62b3f54c8fc12f00 Mon Sep 17 00:00:00 2001
From: localai-org-maint-bot
 <306269227+localai-org-maint-bot@users.noreply.github.com>
Date: Mon, 12 Oct 2026 20:05:51 +0000
Subject: [PATCH] chore(gallery): add Qwen3.8 GSQ-RCO IQ3_S

Complete the existing GSQ-RCO choices with the 3.50-bit build and its
vision projector. Pin downloads and verify both SHA256 values against
Hugging Face LFS metadata and download headers.

Correct the older quality description and document variant selection.

Assisted-by: Codex:gpt-6
---
 docs/content/features/model-gallery.md | 17 ++++++++++
 gallery/index.yaml                     | 46 ++++++++++++++++++++++++--
 2 files changed, 61 insertions(+), 2 deletions(-)

diff --git a/docs/content/features/model-gallery.md b/docs/content/features/model-gallery.md
index a5af4a904..b15745b36 100644
--- a/docs/content/features/model-gallery.md
+++ b/docs/content/features/model-gallery.md
@@ -205,6 +205,23 @@ The files use standard llama.cpp quantization types and the Apache-2.0 license.
 See the [publisher's model card](https://huggingface.co/agentionai/Qwen3.8-27B-AP-GGUF)
 for quantization details. These entries do not enable MTP speculative decoding.
 
+## Qwen3.8-27B GSQ-RCO
+
+The Qwen3.8-27B gallery group includes ISTA DASLab's IQ2_XS, IQ2_S, IQ3_XXS,
+and IQ3_S GGUF builds for llama.cpp. Each includes the BF16 vision projector.
+IQ3_S uses 3.50 bits per weight and has an 11.8 GB weights file.
+
+To select IQ3_S explicitly, run:
+
+```bash
+local-ai models install qwen3.8-27b-q4 --variant qwen3.8-27b-gsq-rco-iq3-s
+```
+
+These builds use the embedded chat template and default to 262,144 context tokens.
+They do not enable MTP speculative decoding.
+See the [publisher's model card](https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF)
+for quantization details and benchmarks.
+
 ## Swift 1.5 Qwen3.8-27B GSQ-RCO
 
 Install `swift-1.5-qwen3.8-27b-gsq-rco` for text chat with llama.cpp.
diff --git a/gallery/index.yaml b/gallery/index.yaml
index 947b4be7d..21846f96c 100644
--- a/gallery/index.yaml
+++ b/gallery/index.yaml
@@ -5843,6 +5843,7 @@
     - 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-gsq-rco-iq3-s
     - model: qwen3.8-27b-q8
   url: github:mudler/LocalAI/gallery/virtual.yaml@master
   urls:
@@ -6381,8 +6382,8 @@
     - 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
+    Qwen3.8-27B with ISTA DASLab's 3.00-bit GSQ-RCO IQ3_XXS mixed
+    quantization. The 10.1 GB model offers a smaller alternative to IQ3_S
     and retains the shared BF16 vision projector.
   last_checked: "2026-08-30"
   overrides:
@@ -6415,6 +6416,47 @@
     - 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-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 recommended 3.50-bit GSQ-RCO IQ3_S mixed
+    quantization. The 11.8 GB model retains the shared BF16 vision projector
+    for image input and uses the embedded chat template with llama.cpp.
+  last_checked: "2026-10-12"
+  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_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-IQ3_S.gguf
+      uri: https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/resolve/d562806dbafae37109975e970aae91b43e73b440/Qwen3.8-27B-GSQ-RCO-IQ3_S.gguf
+      sha256: 64b53b64c7aa39f20a7e54bd80582fe595b1d745624ee8a72e92508c0326d810
+    - filename: llama-cpp/mmproj/qwen3.8-27b-gsq-rco/mmproj-Qwen3.8-27B-BF16.gguf
+      uri: https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/resolve/d562806dbafae37109975e970aae91b43e73b440/mmproj-Qwen3.8-27B-BF16.gguf
+      sha256: 13cb7bebccbd04afc8f4090cb949ecf8937cdf7377c5799b1a0c594e7c0d3e16
 - &spark-x2-5-1-7b
   name: spark-x2.5-1.7b-q4
   url: github:mudler/LocalAI/gallery/virtual.yaml@master