Small 4B model 'Agents-A1-4B' reportedly matches larger agents

Agents-A1-4B is a lightweight, roughly 4-billion-parameter AI model designed to widen what it can handle rather than grow in size. On benchmarks like BrowseComp and GAIA, which test multi-step web research ability, it reportedly scores on par with or above much larger models.

It also shows notable gains on coding and research tasks (SciCode, MLE-Lite, ) and on instruction-following accuracy (IFBench, IFEval) compared to similarly sized small models. Both the and a GGUF (lightweight local-inference format) conversion have been released on .

Key points

  • A roughly 4B-parameter lightweight model claims performance comparable to much larger MoE models
  • Strong results on web-search/reasoning benchmarks like BrowseComp and GAIA
  • Improved scores on coding and research tasks such as SciCode, MLE-Lite, and
  • and a GGUF conversion are published on
Read original