Agents-A1
InternScience · 2026-06-26 · 35.1B parameters
Agents-A1 is InternScience's open-weight 35B-A3B multimodal mixture-of-experts agentic model, post-trained from Qwen3.5-35B-A3B. A full-domain supervised fine-tune is followed by domain teacher models and multi-teacher on-policy distillation across long-horizon search, machine-learning engineering, scientific research, instruction following and tool calling. It has a 262,144-token context and a thinking switch in its chat template, and is released under Apache 2.0.
Benchmark scores
| Benchmark | Score | Source |
|---|---|---|
| BrowseComp | 75.5 | Agents-A1 model card · InternScience |
| Capability | 152.3 | Capability · atlas |
| Frontier Science Research - release-reported score | 40.0 | Agents-A1 model card · InternScience |
| Humanity's Last Exam (with tools) | 47.6 | Agents-A1 model card · InternScience |
| IFBench - release-reported instruction following | 80.6 | Agents-A1 model card · InternScience |
| IFEval - release-reported, metric unspecified | 94.8 | Agents-A1 model card · InternScience |
| LongBench v2 | 60.2 | Agents-A1 model card · InternScience |
| Seal-0 | 56.4 | Agents-A1 model card · InternScience |
| SWE-Shift (Clean) | 14.6 | SWE-Shift paper Table 2 (Agents-A1) · SWE-Shift Authors |
| SWE-Shift (Shifted) | 5.7 | SWE-Shift paper Table 2 (Agents-A1) · SWE-Shift Authors |
| VitaBench | 38.8 | Agents-A1 model card · InternScience |
| τ²-bench (domain average) | 79.8 | Agents-A1 model card · InternScience |