feat(gallery): add Humanizer 12B BF16 variant

Offer the unquantized GGUF alongside the existing Humanizer builds.
Document variant selection and the required completion prompt.

Assisted-by: Codex:gpt-6

docs/content/features/model-gallery.md

@@ -44,6 +44,28 @@ Both views store the view, search, filter, and selection in the URL. Installing
 from Explore does not move you away from the catalog; the entry updates in place
 when the operation finishes.
 
+### Humanizer 12B
+
+[Humanizer 12B](https://huggingface.co/jialinyyzz/humanizer) rewrites English and
+Chinese drafts with the llama.cpp backend. Its variants include quantized GGUF
+builds and an unquantized BF16 build. Each uses an 8,192-token context.
+
+To install BF16 explicitly:
+
+```bash
+local-ai models install humanizer-12b-q4 --variant humanizer-12b-bf16
+```
+
+Omit `--variant` for automatic selection based on available memory. BF16 weights
+require approximately 23.8 GB of disk space, plus memory for inference.
+
+These gallery entries use the completions API. Send the publisher's instruction,
+two newline characters, the trimmed draft, and the separator
+`\n\n### Rewritten:\n\n` as the prompt. Copy the instruction verbatim from
+[the publisher's prompt format](https://huggingface.co/jialinyyzz/humanizer/blob/2e57b6528a4070c33f3e09ca24e3325a84272f68/prompt_format.json).
+Keep the configured sampling settings and do not add stop strings. Split long
+documents at paragraph boundaries and review rewrites for factual changes.
+
 ### MiniCPM-V 4.7 1B
 
 [MiniCPM-V 4.7 1B](https://huggingface.co/openbmb/MiniCPM-V-4.7-1B)

gallery/index.yaml

@@ -3,6 +3,48 @@
 # 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: humanizer-12b-bf16
+  description: |
+    Humanizer is a 12B Gemma 4 fine-tune for rewriting English and Chinese text.
+    This unquantized BF16 GGUF build uses llama.cpp with an 8,192-token context.
+    This entry uses the completions API with the publisher's exact rewriting prompt.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/google/gemma-4-12B
+    - https://huggingface.co/jialinyyzz/humanizer
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - text-generation
+    - rewriting
+    - gemma
+  files:
+    - filename: humanizer-12b-bf16.gguf
+      sha256: 47d79b44c3e15ea2540f4edb63556f2b7e456067d7dd252a42ff103a26a51be9
+      uri: https://huggingface.co/jialinyyzz/humanizer/resolve/2e57b6528a4070c33f3e09ca24e3325a84272f68/humanizer-12b-bf16.gguf
+  overrides:
+    backend: llama-cpp
+    context_size: 8192
+    known_input_modalities:
+      - text
+    known_usecases:
+      - completion
+    mmap: true
+    options:
+      - use_jinja:false
+    parameters:
+      min_p: 0
+      model: humanizer-12b-bf16.gguf
+      repeat_penalty: 1.0
+      temperature: 1.0
+      top_k: 0
+      top_p: 0.95
+    stopwords: []
+    template:
+      completion: '{{.Input}}'
 - url: github:mudler/LocalAI/gallery/virtual.yaml@master
   name: adavar-3b-q4
   description: |
@@ -75532,6 +75574,7 @@
     - model: humanizer-12b-q3-qat
     - model: humanizer-12b-q6
     - model: humanizer-12b-q8
+    - model: humanizer-12b-bf16
 - url: github:mudler/LocalAI/gallery/virtual.yaml@master
   name: humanizer-12b-q6
   description: |

gallery/models/index.yaml

@@ -4,6 +4,56 @@
 # 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: humanizer-12b-bf16
+  name: humanizer-12b-bf16
+  version: 0.0.0-gallery.5a9c8fba6573
+  description: |
+    Humanizer is a 12B Gemma 4 fine-tune for rewriting English and Chinese text.
+    This unquantized BF16 GGUF build uses llama.cpp with an 8,192-token context.
+    This entry uses the completions API with the publisher's exact rewriting prompt.
+  license: apache-2.0
+  urls:
+    - https://huggingface.co/google/gemma-4-12B
+    - https://huggingface.co/jialinyyzz/humanizer
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - text-generation
+    - rewriting
+    - gemma
+  channel: gallery
+  policy:
+    requireSignature: false
+  schemaVersion: 1
+  model:
+    url: github:mudler/LocalAI/gallery/virtual.yaml@master
+    files:
+      - filename: humanizer-12b-bf16.gguf
+        sha256: 47d79b44c3e15ea2540f4edb63556f2b7e456067d7dd252a42ff103a26a51be9
+        uri: https://huggingface.co/jialinyyzz/humanizer/resolve/2e57b6528a4070c33f3e09ca24e3325a84272f68/humanizer-12b-bf16.gguf
+    overrides:
+      backend: llama-cpp
+      context_size: 8192
+      known_input_modalities:
+        - text
+      known_usecases:
+        - completion
+      mmap: true
+      options:
+        - use_jinja:false
+      parameters:
+        min_p: 0
+        model: humanizer-12b-bf16.gguf
+        repeat_penalty: 1.0
+        temperature: 1.0
+        top_k: 0
+        top_p: 0.95
+      stopwords: []
+      template:
+        completion: '{{.Input}}'
 - kind: model
   id: adavar-3b-q4
   name: adavar-3b-q4
@@ -92933,6 +92983,7 @@
       - model: humanizer-12b-q3-qat
       - model: humanizer-12b-q6
       - model: humanizer-12b-q8
+      - model: humanizer-12b-bf16
 - kind: model
   id: humanizer-12b-q6
   name: humanizer-12b-q6