Agents-A1
InternScience/Agents-A1
published Jun 2026 · updated Jul 2026
Agents-A1 is a text-generation model that uses a 35B Mixture-of-Experts architecture to perform long-horizon agentic tasks such as search, engineering, scientific research, instruction following, and tool-calling.
specs
| Task | Text Generation |
| Architecture | Mixture-of-Experts (MoE) |
| Parameters | 35B total (3B active) |
| License | Apache-2.0 |
about this model
Training and Architecture
The model is trained using a three-stage recipe: full-domain supervised fine-tuning for broad agentic alignment, domain-level teacher models for specialized expertise, and multi-teacher domain-routed on-policy distillation with salient vocabulary alignment. This process leverages agentic trajectories averaging 45K tokens in length, built from a long-horizon knowledge-action infrastructure that connects external knowledge, actions, observations, and verifier outcomes.
Key Strengths
- Agentic reasoning: decomposes complex tasks into executable sub-steps and adapts strategies based on intermediate results.
- Tool use: natively supports function calling and integration with APIs, code interpreters, search engines, and other external tools.
- Scientific and professional reasoning: handles tool-integrated scientific reasoning and professional knowledge question answering.
- Instruction following: precisely follows detailed, multi-constraint instructions across diverse domains.
Benchmark Performance
Despite its ~35B parameter class, Agents-A1 achieves overall state-of-the-art results on several challenging benchmarks and remains highly competitive against frontier-scale systems such as GPT-5.5, DeepSeek-V4-pro, and Kimi-K2.6.
| Benchmark | Agents-A1 Score | Notable Comparison |
|---|---|---|
| Seal-0 | 56.4 | Overall SOTA |
| HiPhO | 46.4 | Overall SOTA |
| FrontierScience-Olympiad | 79.0 | Overall SOTA |
| FrontierScience-Research | 40.0 | Overall SOTA |
| IFBench | 80.6 | Overall SOTA |
| IFEval | 94.8 | Overall SOTA |
| BrowseComp | 75.5 | Best among comparable ~35B models |
| XBench-DS-2510 | 86.0 | Best among comparable ~35B models |
| GAIA | 96.0 | Best among comparable ~35B models |
| SciCode | 44.3 | Best among comparable ~35B models |
| HLE with tools | 47.6 | Best among comparable ~35B models |
| MolBench-bind | 56.8 | Best among comparable ~35B models |
The model is released under the Apache-2.0 license. Quantized variants including FP8, Q4_K_M-GGUF, Q8_0-GGUF, and F16-GGUF are available in the Agents-A1 collection on Hugging Face.
best for
- ·Long-horizon search and browsing tasks (e.g., BrowseComp, GAIA)
- ·Scientific research and reasoning with tool integration
- ·Complex instruction following with multiple constraints
- ·Engineering and code generation tasks
FAQ
Agents-A1 is a 35B total parameter Mixture-of-Experts model with approximately 3B active parameters per forward pass.
It is released under the Apache-2.0 license.
Despite being only 35B parameters, Agents-A1 achieves leading or competitive results on benchmarks like Seal-0, IFBench, HiPhO, and FrontierScience-Olympiad, often matching or exceeding trillion-parameter models.
Use the gigarouter OpenAI-compatible endpoint with your API key to send text-generation requests to the model.
It unifies six heterogeneous domains: long-horizon search, engineering tasks, scientific research, instruction following, general agentic tasks, and scientific agentic tasks.
We're benchmarking and onboarding Agents-A1 as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.