autotrust/gemma4-31B-Fable-5-Distilled
autotrust/gemma4-31B-Fable-5-Distilled is a 31.27 billion parameter Gemma-4-based model developed by AutoTrust AI Lab, fine-tuned for agentic coding and tool-use. It achieves a 92.7% pass@1 on HumanEval, a 15.9 point improvement over its base model, while uniquely preserving its multimodal vision capabilities through a layer-freezing LoRA strategy. This model excels at agentic code generation, tool-use planning, and image description, making it suitable for applications requiring both advanced coding and visual reasoning.
Loading preview...
Gemma-4-31B-Fable-5-Distilled: Enhanced Coding with Preserved Vision
Developed by AutoTrust AI Lab, gemma4-31B-Fable-5-Distilled is a 31.27 billion parameter model built upon google/gemma-4-31B-it. It is specifically fine-tuned using agentic coding traces from Fable 5 to significantly boost coding and tool-use performance.
Key Capabilities
- Superior Coding Performance: Achieves an impressive 92.7% pass@1 on HumanEval, marking a +15.9 point improvement over the base
google/gemma-4-31B-itmodel (76.8%). - Preserved Multimodal Vision: Unlike many coding fine-tunes, this model maintains the base model's vision capabilities by applying LoRA adapters only to the upper half of the transformer stack (layers 30-59), leaving the lower multimodal fusion layers (0-29) frozen. This ensures image description quality matches the base model.
- Parameter-Efficient Fine-tuning: The performance lift is achieved with only 0.20% of total parameters trainable (61.2M out of 31.27B), demonstrating highly efficient adaptation.
- Agentic Reasoning: Trained on a small, high-quality dataset of 308 complete tool-use conversations from Fable 5, emphasizing thinking blocks and structured tool calls.
Good for
- Agentic Code Generation: Ideal for tasks requiring chain-of-thought reasoning and code explanation.
- Tool-Use Planning: Capable of generating structured JSON tool-call outputs for integration with external execution layers.
- Multimodal Applications: Excellent for image description and visual reasoning, where preserving vision capabilities alongside coding prowess is crucial.
- General-Purpose Chat: Supports general conversational tasks with an enabled thinking mode for more robust responses.