Hugging Face Overview
Overview of Arch-Function-3B
Katanemo's Arch-Function-3B is a 3.1 billion parameter LLM from the Arch-Function collection, specifically engineered for function calling tasks. Built upon the Qwen 2.5 architecture, it demonstrates high proficiency in interpreting natural language to produce accurate function call outputs, including complex function signatures and parameter identification.
Key Capabilities
- State-of-the-art performance in function calling, achieving competitive results against larger models like GPT-4o on the Berkeley Function-Calling Leaderboard.
- Accurate parameter identification and suggestion, even with ambiguous or incomplete inputs.
- High generalization across various function calling use cases, from API interactions to automated backend tasks.
- Optimized for low-latency, high-throughput performance, suitable for real-time production environments.
- Supports diverse function calling patterns: single, parallel (same function, different parameters), multiple (different functions), and combined parallel & multiple calls.
Performance Highlights
On the Berkeley Function-Calling Leaderboard (BFCL), Arch-Function-3B scored an Overall 57.69%, with strong performance in single-turn tasks (85.19% Non-live AST, 86.18% Non-live Exec, 71.21% Live AST) and high relevance in hallucination metrics (90.24%). It is a core component of the open-source Arch project, an AI-native proxy server for LLMs.
Good for
- Developers requiring robust and accurate function calling capabilities.
- Applications involving automated API interactions and backend task orchestration.
- Integrating LLM-driven tool use in real-time production systems.