gemma4-comfyui-mcp-12b

gemma4-comfyui-mcp-12b (merged bf16 — provider-servable)

Merged full weights of the 12B rung of artokun/gemma4-comfyui-mcp — Gemma 4 12B QLoRA-fine-tuned into a ComfyUI expert that drives the complete comfyui-mcp tool surface (178 tools) in Gemma 4's native tool-call format.

This repo exists for inference providers and self-hosted serving (vLLM / TGI / SGLang need root-layout merged safetensors). For local use, prefer the GGUF ladder: ollama pull artokun/gemma4-comfyui-mcp:12b (also :e4b, :e2b).

• Base: coder3101/gemma-4-12B-it-heretic (Heretic-abliterated Gemma 4 12B)
• Fine-tune: QLoRA r=32/α=32, 2 epochs, trimmed-context tool-menu training on server-verified ComfyUI tool-use trajectories — dataset public at artokun/comfyui-mcp-trajectories
• Format: bf16, 5 sharded safetensors (~24 GB), Gemma 4 unified arch (AutoModelForImageTextToText), trained chat_template.jinja included
• Serving notes: the comfyui-mcp tool payload is large — serve with a generous context window (16K minimum, 64K recommended); temperature 0 recommended for tool precision

Adapter-only, GGUFs, training pipeline, and the full model card live in the main ladder repo.

License: Gemma. Fine-tune by @artokun.