Five Small Open Models in One Binary Reportedly Beat Claude Fable 5 on LiveCodeBench Hard

A team of five 27B-parameter open-weight Qwen models, orchestrated inside a single CLI binary, reportedly scored 49 of 50 on LiveCodeBench Hard — edging out a frontier model and reviving the case for structured multi-agent workflows.

A multi-agent setup built entirely from small open-weight models reportedly outperformed a frontier model on one of the harder coding benchmarks going. According to @VaseGod, a team of five Qwen3.8 27B models, orchestrated inside a single CLI binary, beat Anthropic's Claude Fable 5 on LiveCodeBench Hard. The detail that makes it notable: this isn't a 400-billion-parameter behemoth winning on raw scale. It's five mid-sized models coordinating.

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.