Agents-A1 - An open-source hybrid expert agent model from Shanghai AI Lab

What is Agents-A1?

Agents-A1 is an open-source 35-parameter hybrid expert (MoE) intelligent agent model from the Shanghai AI Lab, designed specifically for complex scientific research tasks. The model constructs a verifiable long-trajectory infrastructure through a knowledge-action graph and employs a three-stage multi-expert distillation training method, overcoming the limitations of large models that rely on massive numbers of parameters for complex reasoning. The model performs comparably to trillion-parameter models in tasks such as long-range search, scientific reasoning, and instruction following.

Main functions of Agents-A1

• Long-range autonomous mission executionIt supports continuous multi-step interaction, environmental feedback, and iterative execution to complete complex scientific research processes.
• Scientific reasoningIt covers multiple stages, including idea generation, solution design, code implementation, and experimental verification.
• Long-range searchIt integrates multiple rounds of information retrieval, cross-validation, and report generation.
• Tool callIt allows for flexible use of external tools and APIs to interact with the environment and dynamically adjust strategies.
• Instructions followed: To complete reasoning, decision-making, and replanning after failure under complex constraints.

Technical Principles of Agents-A1

• Knowledge-Action Graph (KAG)Unlike ordinary knowledge graphs, KAG preserves the complete trajectory of the acquisition, verification, correction and validation of the answer, and continuously expands high-quality long trajectory data through self-game, with an average length of about 45,000 tokens, providing process-level supervision for the model.
• Three-stage training processThe first stage involves full-domain supervised fine-tuning to obtain the basic capabilities of a general intelligent agent. The second stage trains multiple domain expert models for specific tasks, specifically strengthening specialized capabilities such as long-range search, scientific reasoning, and tool invocation. The third stage unifies the capabilities of different experts into a single model through online policy distillation (OPD) of multi-domain routing.
• Distillation of significant vocabulary alignmentThe system employs a multi-teacher domain routing online strategy distillation, which improves the efficiency of cross-domain knowledge transfer through significant vocabulary alignment, and unifies six heterogeneous domains, including search, science, engineering, and instruction compliance, into a single deployable student model.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use Agents-A1

• Environmental preparationClone the GitHub code repository and install Python dependencies, ensuring that the GPU memory is at least 48GB to support 35B MoE model inference.
• Model AcquisitionDownload via HuggingFace CLI or Moda Community command line InternScience/Agents-A1 Complete weights and configuration files.
• Local deployment: Load the model from the local path using the Transformers library, and configure it. device_map="auto" Enables multi-card parallel operation or efficient allocation of a single card.
• Task executionAfter registering external tools such as search APIs and code interpreters in the configuration file, submit research goals in natural language, and the model will automatically decompose the steps, call the tools, and iteratively output the results.

Agents-A1's core advantages

• Small parameters, big capabilitiesWith only 35 parameters, it rivals or surpasses trillion-parameter models on benchmarks such as SEAL-0, IFBench, and HiPhO.
• Interaction Depth Scaling Law: Shift from horizontal expansion of parameter scale to vertical expansion of interaction depth, using a smaller scale to leverage stronger and more complex task processing capabilities.
• General and specialized integrationA single model can uniformly cover multiple fields such as long-range search, scientific reasoning, tool invocation, and instruction compliance.
• Verifiable long trajectoryThe KAG mechanism retains complete action and verification trajectories, ensuring that the process is traceable and the results are verifiable.
• Fully open sourceThe code, models, and technical reports are all open source. Within 10 days of its launch, it has accumulated over 200,000 downloads, and community-derived versions continue to emerge.

Agents-A1 project address

• Project official websitehttps://internscience.github.io/Agents-A1/
• GitHub repositoryhttps://github.com/InternScience/Agents-A1
• HuggingFace model libraryhttps://huggingface.co/InternScience/Agents-A1
• arXiv technical paper: https://arxiv.org/pdf/2606.30616

Comparison of Agents-A1 with similar competing products

Application scenarios of Agents-A1