rautaditya/qwen2.5-0.5B-toolcall-mlx

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 19, 2026Architecture:Transformer Featherless Exclusive Cold

The rautaditya/qwen2.5-0.5B-toolcall-mlx model is a 0.5 billion parameter Qwen2.5-Instruct variant, fine-tuned for tool-calling capabilities on Apple Silicon via MLX. This model is specifically optimized for function calling, enabling it to accurately parse user requests into structured tool calls for functions like `get_weather` and `calculator`. It excels at generating precise tool-call formats and handling argument extraction, making it suitable for agentic workflows and automated task execution.

Loading preview...

Overview

This model, rautaditya/qwen2.5-0.5B-toolcall-mlx, is a specialized version of the Qwen2.5-0.5B-Instruct model, fine-tuned for robust tool-calling functionality. It is designed for efficient deployment on Apple Silicon using the MLX framework, with fused LoRA adapters (rank 16) for optimized inference without needing separate adapter loading.

Key Capabilities

  • Accurate Tool Calling: Demonstrates perfect scores (1.000 / 1.000) on strict 8-case evaluations for tool name and argument extraction, significantly improving over the base model's 0.688 score.
  • Function Schema Adherence: Generates tool calls in a specific JSON format, <tool_call>{"name": ..., "arguments": {...}}</tool_call>, for predefined tools like get_weather and calculator.
  • Argument Handling: Effectively extracts required arguments and prompts for missing ones, avoiding hallucination of values.
  • MLX Optimization: Provided as full fused MLX weights, ensuring native and efficient performance on Apple Silicon.

Performance Highlights

On a 22-case held-out benchmark, the LoRA-tuned model achieved an overall score of 0.864, outperforming the base model's 0.682. It showed perfect scores (1.000) in calculator, clarification, and no-tool scenarios, and strong performance in tool name recognition (0.955).

Good for

  • Agentic Applications: Ideal for developers building AI agents that need to reliably interact with external tools or APIs.
  • Apple Silicon Development: Optimized for MLX, making it a strong choice for local development and deployment on Macs.
  • Structured Output Generation: When precise, schema-compliant JSON output for function calls is critical.

Limitations

  • Limited Vocabulary for Entities: May struggle with unseen or uncommon city spellings due to training data scope.
  • Contextual Ambiguity: Can sometimes over-trigger clarification for ambiguous weather-related phrasing.
  • Non-grounded no_tool answers: Responses for no_tool scenarios are short SFT strings and are not fact-checked or grounded, potentially leading to factual inaccuracies.

Finetunes
1 models