Overview

Arch-Agent-32B is a 32.8 billion parameter large language model developed by katanemo, specifically engineered for advanced function calling and agent-based applications. It is designed to manage sophisticated multi-step and multi-turn workflows, making it highly effective for tasks requiring intelligent tool selection, adaptive planning, and integration with external APIs.

Key Capabilities

• Multi-Turn Function Calling: Maintains context across multiple dialogue turns for evolving tool use.
• Multi-Step Function Calling: Plans and executes sequences of function calls, dynamically adapting based on intermediate results and decomposing complex goals.
• Agentic Capabilities: Provides advanced decision-making and workflow management for complex agentic tasks, including seamless tool coordination and error recovery.

Performance

Arch-Agent-32B demonstrates leading performance on the Berkeley Function-Calling Leaderboard (BFCL), evaluated with a 64K context length using YaRN scaling for multi-turn scenarios. The model is built to deliver reliability and precision across extended function call sequences.

Usage

The model is compatible with the Hugging Face transformers library (version >=4.51.0) and utilizes a specific prompt format to extract JSON output similar to OpenAI's function calling. Example code is provided for quick integration into function calling tasks.