feat(gallery): add Saluki abliterated variants

Add Kujira's mixed IQ2 Saluki fine-tune for llama.cpp with plain and MTP
decoding. Pin the weights and document the text-only entries.

Assisted-by: Codex:gpt-6
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

docs/content/features/model-gallery.md

@@ -61,6 +61,21 @@ Select a specific build with `--variant minicpm-v-4_7-1b-q4`,
 `--variant minicpm-v-4_7-1b-q5`, `--variant minicpm-v-4_7-1b-q6`,
 or `--variant minicpm-v-4_7-1b-q8`.
 
+### Underdog Saluki abliterated builds
+
+`underdog-saluki-27b-1.0-abliterated` provides Kujira's mixed IQ2 adaptation
+of Underdog Saluki for text chat, reasoning, and tool use. Its MTP variant
+enables multi-token prediction with up to two draft tokens. Both builds use
+the same weights, the embedded chat template, and a 32,768-token context.
+
+To select the MTP build explicitly:
+
+```bash
+local-ai models install --variant underdog-saluki-27b-1.0-abliterated-mtp underdog-saluki-27b-1.0-abliterated
+```
+
+These entries are text-only and do not include a vision projector.
+
 ### Repairing an existing model configuration
 
 Gallery template updates do not rewrite YAML files for already installed models.

gallery/index.yaml

@@ -3,6 +3,93 @@
 # 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: underdog-saluki-27b-1.0-abliterated
+  description: |
+    Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed
+    IQ2 weights with an embedded MTP head. This text-only llama.cpp build
+    uses ordinary decoding; the MTP variant enables speculative decoding.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0
+    - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - reasoning
+    - function-calling
+    - uncensored
+  files:
+    - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+      sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc
+      uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+  overrides:
+    backend: llama-cpp
+    context_size: 32768
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    options:
+      - use_jinja:true
+    parameters:
+      model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+      presence_penalty: 1.1
+      temperature: 1.0
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  variants:
+    - model: underdog-saluki-27b-1.0-abliterated-mtp
+- url: github:mudler/LocalAI/gallery/virtual.yaml@master
+  name: underdog-saluki-27b-1.0-abliterated-mtp
+  description: |
+    Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed
+    IQ2 weights with an embedded MTP head. This text-only llama.cpp build
+    enables multi-token prediction with up to two draft tokens.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0
+    - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - reasoning
+    - function-calling
+    - uncensored
+    - mtp
+  files:
+    - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+      sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc
+      uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+  overrides:
+    backend: llama-cpp
+    context_size: 32768
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    options:
+      - use_jinja:true
+      - spec_type:draft-mtp
+      - spec_n_max:2
+    parameters:
+      model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+      presence_penalty: 1.1
+      temperature: 1.0
+      top_k: 20
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
 - url: github:mudler/LocalAI/gallery/virtual.yaml@master
   name: adavar-3b-q4
   description: |

gallery/models/index.yaml

@@ -4,6 +4,109 @@
 # 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: underdog-saluki-27b-1.0-abliterated
+  name: underdog-saluki-27b-1.0-abliterated
+  version: 0.0.0-gallery.cdab61f388f4
+  description: |
+    Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed
+    IQ2 weights with an embedded MTP head. This text-only llama.cpp build
+    uses ordinary decoding; the MTP variant enables speculative decoding.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0
+    - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - reasoning
+    - function-calling
+    - uncensored
+  channel: gallery
+  policy:
+    requireSignature: false
+  schemaVersion: 1
+  model:
+    url: github:mudler/LocalAI/gallery/virtual.yaml@master
+    files:
+      - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+        sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc
+        uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+    overrides:
+      backend: llama-cpp
+      context_size: 32768
+      function:
+        automatic_tool_parsing_fallback: true
+        grammar:
+          disable: true
+      known_usecases:
+        - chat
+      options:
+        - use_jinja:true
+      parameters:
+        model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+        presence_penalty: 1.1
+        temperature: 1.0
+        top_k: 20
+        top_p: 0.95
+      template:
+        use_tokenizer_template: true
+    variants:
+      - model: underdog-saluki-27b-1.0-abliterated-mtp
+- kind: model
+  id: underdog-saluki-27b-1.0-abliterated-mtp
+  name: underdog-saluki-27b-1.0-abliterated-mtp
+  version: 0.0.0-gallery.774b2b47f0dc
+  description: |
+    Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed
+    IQ2 weights with an embedded MTP head. This text-only llama.cpp build
+    enables multi-token prediction with up to two draft tokens.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0
+    - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - reasoning
+    - function-calling
+    - uncensored
+    - mtp
+  channel: gallery
+  policy:
+    requireSignature: false
+  schemaVersion: 1
+  model:
+    url: github:mudler/LocalAI/gallery/virtual.yaml@master
+    files:
+      - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+        sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc
+        uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+    overrides:
+      backend: llama-cpp
+      context_size: 32768
+      function:
+        automatic_tool_parsing_fallback: true
+        grammar:
+          disable: true
+      known_usecases:
+        - chat
+      options:
+        - use_jinja:true
+        - spec_type:draft-mtp
+        - spec_n_max:2
+      parameters:
+        model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf
+        presence_penalty: 1.1
+        temperature: 1.0
+        top_k: 20
+        top_p: 0.95
+      template:
+        use_tokenizer_template: true
 - kind: model
   id: adavar-3b-q4
   name: adavar-3b-q4