nakue/qwen2.5-0.5b-funccall

TEXT GENERATIONConcurrency Cost:1Model Size:0.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Jun 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

nakue/qwen2.5-0.5b-funccall is a 0.5 billion parameter Qwen2.5-Instruct model fine-tuned by nakue for efficient function calling. It acts as a specialized router, taking natural language queries and tool schemas to output precise JSON function calls. This model is optimized for agent loops and CLI dispatchers, enabling structured function execution without relying on larger, more expensive general-purpose LLMs.

Loading preview...

Overview

nakue/qwen2.5-0.5b-funccall is a specialized 0.5 billion parameter model, fine-tuned from unsloth/Qwen2.5-0.5B-Instruct using LoRA. Its primary purpose is to serve as a function-calling router, translating natural language requests into structured JSON function calls. This model is designed to be a lightweight and accurate alternative for scenarios where calling a large general-purpose model for every function dispatch is cost-prohibitive.

Key Capabilities

  • Efficient Function Calling: Takes a user query and a set of tool schemas, then outputs clean, parseable JSON representing the correct function call(s).
  • Compact Size: At 0.5B parameters, it offers a significantly smaller footprint compared to larger LLMs, making it suitable for resource-constrained environments or high-throughput applications.
  • Specialized Training: Fine-tuned on the Salesforce/xlam-function-calling-60k dataset, which includes 60,000 verified function-calling examples.
  • Direct JSON Output: Generates only the JSON array of {"name": ..., "arguments": ...} objects, without additional prose or markdown fences, simplifying parsing.

Intended Use Cases

  • Agent Loops: Ideal for dispatching tools within AI agent architectures.
  • CLI Dispatchers: Mapping natural language commands to specific command-line functions.
  • Structured Tool Use: Any system requiring a reliable and cost-effective way to convert user requests into structured function invocations.

Limitations

  • Currently trained and tested only on simple and multiple-style single-turn function calls; not designed for multi-turn conversations or parallel calls.
  • Output reliability can be sensitive to tool-schema formatting, especially with large or unusual schemas outside the training distribution.
  • Evaluation against established benchmarks like the Berkeley Function-Calling Leaderboard (BFCL) is still pending. Users should note that performance claims against larger models are not yet verified.

Finetunes
1 models