Gemma 4 12B V2 Coding Agentic

Gemma4-12B v2 (yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF) is a text-generation model specialized for coding and agentic tool-use tasks, fine-tuned from Google's Gemma 4 12B instruction-tuned base model. It is designed to read, reason, use tools, and work through multi-step technical tasks before acting, making it suitable for local, private agentic workflows.

Key Strengths and Benchmarks

The model achieves roughly 3.5× higher performance than the base model on the tau2-bench telecom benchmark, which mirrors real terminal/debugging work by requiring diagnose → fix → verify loops. In a separate fabrication probe, the model grounds actions before acting (0% fabrication, on par with the base model). The base model tends to abandon tasks (transfer to human), while v2 persists through the agentic loop.

Trade-offs and Specialization

This fine-tune is specialized for coding, terminal, and technical-agentic work. It trades a small amount of general-knowledge breadth (MMLU-Pro scores slightly below the base model) for its agentic capabilities. It is not designed for customer-service tasks (e.g., tau2-bench retail).

Training and Architecture

v2 builds on v1 with a significant agentic push: multi-step tool-use trajectories (read → reason → act → verify) in Gemma 4's native tool protocol, verified chain-of-thought for Python coding tasks, and curated reasoning/instruction data to maintain broad competence. All reasoning is distilled CoT from Opus 4.8. The model uses Gemma's native thought channel before answering, and is distributed under Apache 2.0.