🤖 Gemma-4-12B Agentic Coder (v2) — GGUF

This collection hosts highly optimized GGUF formats of the fine-tuned Gemma-4-12B-it model, tailored specifically for local coding assistance and autonomous multi-step agent workflows.

The primary goal of this fine-tune is to bridge the gap between heavy cloud models and consumer-grade hardware. It has been strictly optimized to excel at tool-use, complex multi-step reasoning, and terminal-based task execution while remaining lightweight enough to run locally.

📚 What’s new in v2 (training)

v2 builds on the v1 coder by introducing an updated agentic workflow:

• 🛠️ Agentic / terminal — Multi-step tool-use trajectories (read → reason → act → verify) utilizing the native tool protocol. This update addresses the single-step limitation observed in the previous version.
• 💻 Coding — Verified chain-of-thought (CoT) processing for Python tasks, requiring passing test results, supplemented by the Fable-5-redo dataset for more complex cases.
• 📚 General — A selected subset of reasoning and instruction data designed to maintain general competence.

The reasoning model utilizes distilled CoT, with traces reconstructed using the specified workflows.

📦 Quantization Details (GGUF)

We offer several quantization levels to accommodate different hardware constraints.

📝 Base Model & Licensing

• Quantized Model: yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2
• Original Base Model: google/gemma-4-12B-it
• License: Gemma Terms of Use (Permissible for modification, redistribution, and commercial use under Google’s official Gemma License Agreement).