Nexusflow/Athene-V2-Agent

Athene-V2-Agent: Advanced Tool Use and Agentic Capabilities

Athene-V2-Agent, developed by Nexusflow, is an open-source agent LLM built upon the Qwen-2.5-72B-Instruct model, featuring 72.7 billion parameters. It is specifically engineered to surpass existing models, including GPT-4o, in complex function calling and agentic tasks.

Key Capabilities and Performance

• Superior Agentic Performance: The model demonstrates an 18% higher success rate in single function calls and a 17% higher success rate in agentic tasks compared to GPT-4o.
• Complex Reasoning and Planning: Athene-V2-Agent is capable of reasoning and planning for trajectories that require multiple, deeply nested tool calls to resolve a single query.
• Generalization: It exhibits strong generalization capabilities, performing effectively even with functions and agentic settings it was not explicitly trained on.
• OpenAI API Compatibility: Designed for seamless integration, it can be used as a drop-in replacement in any OpenAI API-compatible environment via a custom VLLM Docker image.

Usage Recommendations

• Custom Prompting: For optimal performance, Nexusflow strongly recommends using their custom VLLM Docker image, as the model's unique prompting style for executable calls is baked into this environment.
• Docstring Quality: Providing well-indented, detailed, and well-written docstrings for tools significantly enhances accuracy.
• Sampling Settings: It is recommended to set sampling to False and use a zero temperature for consistent and controllable behavior.
• Controllable Behavior: The model is highly tunable for system integration, allowing for explicit control over behaviors like rejecting irrelevant queries (using a no_relevant_function tool) or engaging in chat (using a chat tool).