From a53dc001f368f67a73fae501933ff5789ea69356 Mon Sep 17 00:00:00 2001
From: localai-org-maint-bot
 <306269227+localai-org-maint-bot@users.noreply.github.com>
Date: Wed, 9 Sep 2026 12:06:22 +0000
Subject: [PATCH] feat(gallery): add Dirk Q5 and Q6 variants

Offer intermediate GGUF builds with MTP and the shared vision projector.
Pin downloads and document explicit variant selection.

Assisted-by: Codex:GPT-6
---
 docs/content/features/model-gallery.md | 10 ++++
 gallery/index.yaml                     | 82 +++++++++++++++++++++++++-
 2 files changed, 91 insertions(+), 1 deletion(-)

diff --git a/docs/content/features/model-gallery.md b/docs/content/features/model-gallery.md
index b8cf63898..63a00a583 100644
--- a/docs/content/features/model-gallery.md
+++ b/docs/content/features/model-gallery.md
@@ -68,6 +68,16 @@ GALLERIES=[{"name":"<GALLERY_NAME>", "url":"<GALLERY_URL"}]
 
 The models in the gallery will be automatically indexed and available for installation.
 
+## Dirk quantization variants
+
+Dirk Qwen3.8 27B offers Q4_K_XL, Q5_K_XL, Q6_K_XL, and Q8_K_XL GGUF builds
+with llama.cpp, MTP speculative decoding, and the shared F16 vision projector.
+Select Q6 explicitly with:
+
+```bash
+local-ai models install dirk-qwen3.8-27b-q4 --variant dirk-qwen3.8-27b-q6
+```
+
 ## Gallery mirrors
 
 A gallery entry can declare a `mirrors` list of alternative locations for the same index file. Mirrors exist for availability, not for load balancing: LocalAI always prefers the `url`, and only falls back to the mirrors, in the order you listed them, when the one before it cannot be fetched. If the primary works, the mirrors are never contacted.
diff --git a/gallery/index.yaml b/gallery/index.yaml
index 804049539..33260db6f 100644
--- a/gallery/index.yaml
+++ b/gallery/index.yaml
@@ -1441,6 +1441,8 @@
 - &dirk-qwen3-8-27b
   name: "dirk-qwen3.8-27b-q4"
   variants:
+    - model: dirk-qwen3.8-27b-q5
+    - model: dirk-qwen3.8-27b-q6
     - model: dirk-qwen3.8-27b-q8
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -1453,7 +1455,7 @@
     262K-token context window.
 
     This default entry uses the Q4_K_XL GGUF and F16 vision projector. A
-    higher-quality Q8_K_XL build is available as a variant.
+    choice of Q5_K_XL, Q6_K_XL, and Q8_K_XL builds is available through variants.
   license: "apache-2.0"
   tags:
     - llm
@@ -1541,6 +1543,84 @@
     - filename: llama-cpp/mmproj/dirk-qwen3.8-27b/mmproj-F16.gguf
       uri: huggingface://peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF/mmproj-F16.gguf
       sha256: cbb841a9ee0636b2ec172f5bb8df2ea8dfeb01e90fe7c6126581d662a0b4e43e
+- !!merge <<: *dirk-qwen3-8-27b
+  name: "dirk-qwen3.8-27b-q5"
+  variants: []
+  last_checked: "2026-09-09"
+  description: |
+    Dirk in the higher-quality Q5_K_XL GGUF format, with MTP speculative
+    decoding and the shared F16 vision projector for multimodal prompts.
+  overrides:
+    backend: llama-cpp
+    context_size: 262144
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+      - vision
+    mmproj: llama-cpp/mmproj/dirk-qwen3.8-27b/mmproj-F16.gguf
+    options:
+      - use_jinja:true
+      - spec_type:draft-mtp
+      - spec_n_max:6
+      - spec_p_min:0.75
+    parameters:
+      min_p: 0
+      model: llama-cpp/models/dirk-qwen3.8-27b/Dirk-Qwen3.8-27B-UD-Q5_K_XL.gguf
+      repeat_penalty: 1
+      temperature: 0.6
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/dirk-qwen3.8-27b/Dirk-Qwen3.8-27B-UD-Q5_K_XL.gguf
+      uri: https://huggingface.co/peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF/resolve/52cb3e759635ab4605e08790b6c47df8adcf0744/Dirk-Qwen3.8-27B-UD-Q5_K_XL.gguf
+      sha256: 43d3b23be6e21ddc5f535fc78791ccbbe890c9cafb542a7d8ec5535d171bca3b
+    - filename: llama-cpp/mmproj/dirk-qwen3.8-27b/mmproj-F16.gguf
+      uri: https://huggingface.co/peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF/resolve/52cb3e759635ab4605e08790b6c47df8adcf0744/mmproj-F16.gguf
+      sha256: cbb841a9ee0636b2ec172f5bb8df2ea8dfeb01e90fe7c6126581d662a0b4e43e
+- !!merge <<: *dirk-qwen3-8-27b
+  name: "dirk-qwen3.8-27b-q6"
+  variants: []
+  last_checked: "2026-09-09"
+  description: |
+    Dirk in the higher-quality Q6_K_XL GGUF format, with MTP speculative
+    decoding and the shared F16 vision projector for multimodal prompts.
+  overrides:
+    backend: llama-cpp
+    context_size: 262144
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+      - vision
+    mmproj: llama-cpp/mmproj/dirk-qwen3.8-27b/mmproj-F16.gguf
+    options:
+      - use_jinja:true
+      - spec_type:draft-mtp
+      - spec_n_max:6
+      - spec_p_min:0.75
+    parameters:
+      min_p: 0
+      model: llama-cpp/models/dirk-qwen3.8-27b/Dirk-Qwen3.8-27B-UD-Q6_K_XL.gguf
+      repeat_penalty: 1
+      temperature: 0.6
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/dirk-qwen3.8-27b/Dirk-Qwen3.8-27B-UD-Q6_K_XL.gguf
+      uri: https://huggingface.co/peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF/resolve/52cb3e759635ab4605e08790b6c47df8adcf0744/Dirk-Qwen3.8-27B-UD-Q6_K_XL.gguf
+      sha256: 3ea8eebc1ec4ef820da97cfac526ddd1eed0872d4c8b8db869f2c78a79506c42
+    - filename: llama-cpp/mmproj/dirk-qwen3.8-27b/mmproj-F16.gguf
+      uri: https://huggingface.co/peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF/resolve/52cb3e759635ab4605e08790b6c47df8adcf0744/mmproj-F16.gguf
+      sha256: cbb841a9ee0636b2ec172f5bb8df2ea8dfeb01e90fe7c6126581d662a0b4e43e
 - name: "qwen3.8-27b-dflash2"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls: