InternScience/Agents-A1

What it solves

Agents-A1 addresses the challenge of achieving frontier-level AI performance (comparable to trillion-parameter models) without requiring massive parameter counts. It focuses on "scaling the agent horizon"—improving the model's ability to handle long-term trajectories and diverse specialized abilities—to enable a 35B parameter model to compete with much larger systems in complex, multi-step tasks.

How it works

The project utilizes a 35B Mixture-of-Experts (MoE) architecture trained through a three-stage process:

1. Full-domain SFT: Aligns the base model with general agentic behaviors.
2. Domain-level Teacher Training: Creates specialized teacher models for specific domains of expertise.
3. Multi-teacher Distillation: Uses domain-routed on-policy distillation with salient vocabulary alignment to merge six heterogeneous domains into a single student model.

To support this, the team built a knowledge-action infrastructure that connects external knowledge, actions, and observations, allowing the model to train on trajectories averaging 45K tokens.

Who it’s for

• Developers and Enterprises: Looking to integrate a highly capable agentic model into their applications.
• AI Researchers: Interested in agent-horizon scaling and efficient knowledge transfer via distillation.
• Local AI Enthusiasts: Who can leverage the 4B version or quantized variants for local deployment on hardware like Macs.

Highlights

• Agentic Reasoning: Capable of decomposing complex tasks into executable sub-steps and adapting strategies based on results.
• Tool Integration: Native support for function calling, APIs, code interpreters, and search engines.
• Long-Context Handling: Strong coherence and recall for extended conversations and documents.
• High Efficiency: Delivers competitive performance against frontier models (like GPT-5.5) while remaining in the 35B parameter class.