Agents-A1 is a 35B Mixture-of-Experts model purpose-built for agentic workflows that require planning, tool use, and sustained reasoning over long inputs. Rather than scaling parameter count alone, its design philosophy emphasizes scaling the task horizon, enabling the model to handle long-horizon search, engineering work, scientific research, instruction following, and structured tool-calling scenarios. The model supports a served context length of 256K, making it well suited for extended documents, multi-turn agent sessions, and workflows where intermediate state must be preserved across many steps. As an open-weight release, the full model is available on Hugging Face and ModelScope, with inference compatibility for Hugging Face Transformers, vLLM, and SGLang, giving developers flexible deployment options for local or hosted agent systems.
In practical use, Agents-A1 focuses on three capability areas that map directly to real agent pipelines: decomposing complex goals into executable sub-steps and adapting based on intermediate observations, calling external tools such as APIs, code interpreters, and search engines, and maintaining coherence and recall across long contexts. The model is evaluated across six directions, including long-horizon search, engineering tasks, scientific research, instruction following, general agentic tasks, and scientific agentic tasks, positioning it as a strong foundation for research assistants, automated engineering workflows, and tool-augmented assistants. Subsequent ecosystem releases include quantized variants and a smaller 4B sibling model aimed at local AI assistant use, extending the lineup for resource-constrained deployments while preserving the agent-oriented design.
