InternScience releases Agents-A1-4B, a 4B-parameter model for long-horizon agentic tasks

InternScience has released Agents-A1-4B, a dense model based on Qwen3.5-4B designed to improve performance in long-horizon search and agentic workflows without increasing parameter count.

• The model achieves 75.5 on BrowseComp and 90.0 on XBench-DS-2510, outperforming the base Qwen3.5-4B baseline of 47.2 and 73.0 respectively.
• It scores 96.0 on GAIA and 56.4 on Seal0, setting new state-of-the-art results for models in its size class.
• Agents-A1-4B demonstrates significant gains in engineering tasks, reaching 44.3 on SciCode and 43.9 on MLE-Lite.
• The release includes GGUF checkpoints available via Hugging Face alongside a technical report detailing the scaling methodology.