Gemma 4 12B Agentic Fable 5 Compose: A Local Coding Model

Model Performance and Specialization

Gemma 4 12B Agentic Fable 5 Compose is a fine-tuned iteration of Google's Gemma 4 12B Instruct model, specifically optimized for coding, terminal workflows, and multi-step agentic tasks. While the model demonstrates a claimed 3.5x improvement over the base model on the tau2-bench telecom test, scoring 55% compared to the base model's 15%, it sacrifices general-purpose knowledge. It performs lower than the base model on MMLU Pro benchmarks, making it a specialized tool rather than a general-purpose chatbot.

Implementation and Configuration

To achieve stable output, the model requires specific sampling parameters to prevent repetition and token leakage. Users should configure their inference engine with the following settings:

• Temperature: 1
• Top P: 0.95
• Top K: 64
• Repetition Penalty: 1.1

For local deployment, the Q4 K M quantization is the recommended balance between memory footprint (approximately 6.87 GB) and performance. The model can be served via Ollama or LM Studio, which provides an OpenAI-compatible API server for integration with editors like Zed. When using Zed, users should manually adjust the context length from the default 4,096 tokens to 8,122 tokens to accommodate more complex agentic workflows.

Reliability and Practical Use

Despite its capability in one-shot coding tasks and refactoring, the model is currently prone to bugs, including broken output, repetition, and the exposure of raw tool-call tokens. These issues are exacerbated during multi-step agentic loops where the model must manage file inspection and iterative tool execution. It is not currently recommended for production-grade repository management or tasks requiring high reliability, though it remains a viable candidate for local experimentation and lightweight coding assistance.