Shanghai AI Laboratory · 2026-06-30 · notable
Agents-A1 — Shanghai AI Lab 35B MoE matches trillion-parameter agents
Agents-A1 is an open-weight 35B Mixture-of-Experts agent from Shanghai AI Laboratory. It posts SOTA scores on SEAL-0 (56.4), FrontierScience-Research (40.0), and IFBench (80.6), and the paper claims parity with trillion-parameter agents.
Open-weight 35B agent from Shanghai AI Lab posts SOTA on SEAL-0, IFBench, and FrontierScience-Research.
Key specs
| Parameters | 35B MoE |
|---|---|
| Seal 0 | 56.4 (SOTA) |
| Ifbench | 80.6 (SOTA) |
What is it?
Agents-A1 is a 35B Mixture-of-Experts agent model from Shanghai AI Laboratory's InternScience group, released Apache-2.0 on Hugging Face. The model targets long-horizon search, scientific research, software engineering, instruction following, and tool calling, and uses Qwen3.5-35B-A3B as its base.
How does it work?
Training runs in three stages: full-domain supervised fine-tuning to align broad agentic behavior, a set of domain-level teacher models that each specialize in one skill (search, code, tool use), and multi-teacher on-policy distillation that merges all the teachers back into one student. The infrastructure handles agentic trajectories averaging 45K tokens with knowledge, action, observation, and verification traces.
Why does it matter?
Agents-A1 hits SOTA on SEAL-0 (56.36 for long-horizon search), IFBench (80.61 for instruction following), and FrontierScience-Research (40.0), and the paper says it reaches parity with trillion-parameter systems like Kimi K2.6 and DeepSeek V4-Pro. That puts a frontier-class agent on a single 8-GPU node, with Apache-2.0 weights anyone can fine-tune or run locally.
Who is it for?
Open-source agent builders, scientific-research labs, anyone who needs a SOTA agent without paying frontier API rates.
Try it
huggingface.co/InternScience/Agents-A1