Agents-A1: 35B Mixture‑of‑Experts Agent Model for Long‑Horizon, Tool‑Enabled Reasoning

Agents‑A1 is a 35‑billion‑parameter Mixture‑of‑Experts (MoE) model released by InternScience. Built on the Qwen3.5‑MoE architecture and packaged for the Hugging Face Transformers library, it targets the *text‑generation* pipeline while offering native support for function calling and tool integration. The model is released under an Apache‑2.0 license and is compatible with popular serving frameworks such as vLLM and SGLang, allowing deployments with up to 262k token context windows.

The model is explicitly engineered for heterogeneous agentic abilities across five domains: long‑horizon search, engineering, scientific research, instruction following, and tool‑calling. Training follows a three‑stage paradigm—full‑domain supervised fine‑tuning, domain‑specific teacher models, and multi‑teacher on‑policy distillation—enabling the model to decompose complex tasks, plan ahead, and adapt based on intermediate observations. Benchmarks reported in the technical report (arXiv:2606.30616) show Agents‑A1 achieving state‑of‑the‑art results on a range of agentic and scientific tasks, often outperforming comparable ~35B models and narrowing the gap to much larger systems.

The repository provides model weights, configuration files, and an open‑source evaluation framework for reproducible agent capability testing. Detailed usage instructions cover both SGLang and vLLM, with recommended sampling parameters (temperature 0.85, top_p 0.95, etc.) and examples for standard, tool‑use, and language‑model‑only deployments. Quantized variants are also available through community collections, making the model accessible on a variety of hardware platforms, including Macs via the mlx‑community.

Overall, Agents‑A1 represents a highly capable, general‑purpose agentic model that can be directly integrated into applications requiring multi‑step reasoning, tool interaction, and domain‑specific expertise without the overhead of trillion‑parameter models.