Summary

Agents-A1 is a 35B parameter Mixture-of-Experts agentic model achieving trillion-parameter-level performance by scaling the agent horizon. It targets developers and enterprises seeking capable, efficient AI agents excelling in complex task decomposition, tool integration, and long-context understanding, significantly narrowing the gap with larger frontier models.

How It Works

The core innovation scales agent horizon via extended trajectories (average 45K tokens) and heterogeneous agent abilities. This is supported by a long-horizon knowledge-action infrastructure integrating external knowledge, actions, observations, and verifiers. Training employs a three-stage recipe: full-domain supervised fine-tuning, domain-level teacher model training, and multi-teacher domain-routed on-policy distillation with salient vocabulary alignment, unifying diverse expertise into one student model.

Quick Start & Requirements

Installation involves Python 3.12 environment setup via uv and installing serving frameworks like SGLang (pip install sglang) or vLLM (pip install vllm --torch-backend=auto). The model supports a 262,144 token context length. Serving commands are provided for SGLang and vLLM to launch API endpoints. A link to the technical report (arXiv:2606.30616) is available.

Highlighted Details

• Performance: Delivers highly competitive performance against frontier models like GPT-5.5 and DeepSeek-V4-pro, despite its ~35B parameter class.
• Benchmark Achievements: Achieves State-of-the-Art (SOTA) on Seal-0, HiPhO, FrontierScience-Olympiad, FrontierScience-Research, IFBench, and IFEval. Ranks best among comparable models on BrowseComp, XBench-DS-2510, GAIA, SciCode, HLE with tools, and MolBench-bind.
• Key Capabilities: Excels in Agentic Reasoning (task decomposition, planning, adaptation), native Tool Use (function calling), Long-Context Understanding, and precise Instruction Following across diverse domains.
• Model Variants: Open-sourced quantized model variants, including those runnable on Mac via mlx-community.

Maintenance & Community

The project welcomes feedback from developers and enterprises. Specific community channels (e.g., Discord, Slack), roadmap details, or notable contributors beyond the author list in the citation are not detailed in the provided README.

Licensing & Compatibility

The provided README does not specify a license type. This absence requires further investigation for commercial use or closed-source integration compatibility.

Limitations & Caveats

No explicit limitations are detailed. The README focuses on capabilities and performance. Serving the model likely requires significant computational resources, though specific hardware requirements beyond GPU are not itemized.