4B dense
262,144 tokens
Apache 2.0
Official small-model release in the Agents-A1 family from InternScience, positioned as a long-horizon agentic model that brings the series' search, engineering, scientific reasoning, instruction-following, and tool-use behavior into a locally deployable 4B dense checkpoint. The public Agents-A1 paper focuses on the flagship 35B-A3B model and frames the family around scaling agent horizon rather than raw parameter count, using long-horizon knowledge-action trajectories with actions, observations, and verifier outcomes plus multi-teacher domain-routed on-policy distillation across heterogeneous agent domains. The 4B release, announced separately on July 14, 2026, is presented in the model cards as a compact dense variant that still performs strongly on long-horizon search, engineering and research tasks, instruction following, and scientific agentic benchmarks. Official benchmark tables claim clear gains over similarly sized models such as Qwen3.5-4B on BrowseComp, XBench-DS-2510, GAIA, FrontierScience-Research, and IFEval, with some results approaching or surpassing larger MoE models like Nex-N2-mini and Qwen3.6 on selected agentic tasks. The public release also emphasizes native function calling and tool integration, support for SGLang and vLLM with OpenAI-compatible serving, and multimodal capability in the FP8 variant through image-text usage examples. InternScience positions the model as a practical small open agent for local assistants and long-horizon workflows, while tying it back to the broader Agents-A1 training recipe that aims to reach trillion-parameter-level agent performance by scaling trajectories and heterogeneous abilities rather than scaling parameters alone.