Agent-Ark/Toucan-Qwen2.5-32B-Instruct-v0.1

Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Oct 2, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

Agent-Ark/Toucan-Qwen2.5-32B-Instruct-v0.1 is a 32.8 billion parameter instruction-tuned model based on Qwen2.5, developed by Agent-Ark. It is specifically fine-tuned on the Toucan-1.5M dataset, a large-scale synthetic dataset of over 1.5 million tool-agent trajectories, to enhance tool use capabilities in agentic LLMs. This model excels at complex, multi-turn, and parallel tool-use tasks, outperforming larger closed-source models on tool-use benchmarks. It is designed for applications requiring advanced agentic reasoning and tool interaction.

Loading preview...

Agent-Ark/Toucan-Qwen2.5-32B-Instruct-v0.1: Advanced Tool-Use Agent

This model is a 32.8 billion parameter instruction-tuned variant of Qwen2.5, developed by Agent-Ark. It has been fine-tuned on a curated subset of the Toucan-1.5M dataset, which is the largest fully synthetic tool-agent dataset to date, comprising over 1.5 million trajectories. These trajectories are synthesized from 495 real-world Model Context Protocols (MCPs) across 2,000+ tools, focusing on diverse, realistic, and challenging multi-tool tasks.

Key Capabilities

  • Enhanced Tool Use: Specifically optimized for complex tool interaction, including multi-round, multi-turn, sequential, and parallel tool calls.
  • Agentic Reasoning: Designed to improve the agentic capabilities of LLMs by leveraging a dataset built from authentic MCP environments.
  • Superior Performance: Models fine-tuned on Toucan-1.5M have demonstrated performance improvements, outperforming larger closed-source counterparts on the BFCL V3 benchmark and extending the Pareto frontier on the MCP-Universe benchmark.
  • Extensive Training Data: Fine-tuned on 119.3K instances from the Toucan-1.5M dataset, including extensions for irrelevance, diversification, and multi-turn interactions.

Good For

  • Developing advanced AI agents that require sophisticated tool-use abilities.
  • Applications involving complex, multi-step problem-solving with external tools.
  • Research and development in agentic LLMs and tool-augmented language models.

For detailed methodology and training specifics, refer to the technical report.

Finetunes
1 models