norecyc/lastbox-gemma4-e2b-v6-toolprior

VISIONConcurrency Cost:1Model Size:5.1BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 19, 2026License:gemmaArchitecture:Transformer Cold

The norecyc/lastbox-gemma4-e2b-v6-toolprior is a 5.1 billion parameter Gemma-4 based language model developed by norecyc, specifically fine-tuned for enhanced tool emission and agentic behavior. This model achieves a 72% tool emission rate and a 0.608 agentic score, significantly improving over its base model. It is optimized for deployment on resource-constrained devices like the Raspberry Pi 5, making it suitable for edge AI applications requiring robust tool use.

Loading preview...

LastBox v6: Enhanced Tool Emission for Edge Devices

The norecyc/lastbox-gemma4-e2b-v6-toolprior is a 5.1 billion parameter Gemma-4 based model, specifically engineered to dramatically improve tool emission and agentic performance. Developed by norecyc, this version represents a significant leap from its predecessor, lastbox-gemma4-e2b-sft-v3, achieved through a targeted 12-minute SFT (Supervised Fine-Tuning) pass on tool-only pairs.

Key Capabilities & Improvements

  • High Tool Emission Rate: Achieves a 72% tool emission rate, a substantial increase from near 0% in the base model.
  • Agentic Score: Demonstrates a 38x improvement in agentic score, reaching 0.608.
  • Robust Tool Accuracy: Features 64% tool accuracy and 56% argument validity.
  • Optimized for Edge Deployment: Designed for efficient operation on devices like the Raspberry Pi 5, utilizing Q4_K_M quantization.
  • Targeted SFT Approach: The model's success stems from a focused SFT warmup on 1,034 [user, assistant_tool_call] pairs, setting a strong prior for first-turn tool emission.
  • System Prompt Dependency: To achieve the stated tool emission rates, the model requires the full training-time system prompt, including a 7-tool definitions JSON block, at inference.

When to Use This Model

  • Agentic Applications: Ideal for scenarios requiring reliable tool use and agentic behavior in language models.
  • Resource-Constrained Environments: Excellent choice for deployment on edge devices such as the Raspberry Pi 5 due to its optimized quantization.
  • Tool-Calling Tasks: Particularly suited for tasks where the model needs to accurately identify and call external tools based on user prompts.