A 3-Billion Parameter Agent Model Is Outperforming Giants on SWE-bench — Meet ROME
Chinese researchers released ROME, a 3B-parameter model that achieves state-of-the-art agentic performance on SWE-bench using a novel chunk-level reinforcement learning approach, challenging the assumption that capable agents require massive models.
A team of Chinese researchers released ROME, a 3-billion parameter model that, as @simplifyinAI put it, "breaks every known scaling law for agents." The model achieves competitive or superior performance on SWE-bench — the standard benchmark for AI software engineering agents — compared to models orders of magnitude larger.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.