Jables/qwen2.5-0.5b-toolcall-v2
Jables/qwen2.5-0.5b-toolcall-v2 is a 0.5 billion parameter Qwen2.5-Instruct fine-tune specifically designed for in-browser tool-calling. It generates small, strict JSON decisions for routing tasks within web applications like The Daily Synthesis editor. Optimized for client-side deployment via Transformers.js, this model excels at structured output generation for signal, trend, paper, and connection tasks, offering significantly improved accuracy and reduced malformed outputs compared to its predecessor.
Loading preview...
Overview
Jables/qwen2.5-0.5b-toolcall-v2 is a specialized 0.5 billion parameter model, fine-tuned from Qwen2.5-0.5B-Instruct, to perform in-browser tool-calling. Its primary function is to emit small, strict JSON decisions for routing various tasks (signal, trend, paper, connection) within web applications, such as The Daily Synthesis editor (ZeroClaw). This model is designed for client-side deployment using Transformers.js.
Key Capabilities
- Tool-Call Routing: Generates structured JSON outputs for specific routing decisions.
- In-Browser Execution: Optimized for efficient client-side inference via Transformers.js.
- Improved Accuracy: Achieves significantly higher accuracy (66.9% for
q4f16ONNX) and lower malformed output rates (2.8%) on a 181-case router benchmark compared to previous iterations. - Quantization Awareness: Ships with carefully selected weight-only 4-bit ONNX quantizations (
q4f16for WebGPU,q4for WASM/CPU) to preserve accuracy, as standard 8-bit dynamic quantization severely degrades performance for this model.
Good for
- Client-side AI applications: Ideal for web-based tools requiring lightweight, fast, and accurate structured output generation.
- Tool-use and function calling: Specifically fine-tuned for scenarios where a model needs to decide which tool or function to invoke based on user input.
- Resource-constrained environments: Its small size (0.5B parameters) and optimized quantization make it suitable for deployment where computational resources are limited, such as in a browser.