168 points92 comments
Five Qwen3.8-27B agents outcode Fable 5 on LiveCodeBench
What is Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard?
A Memdoor workflow spins up five specialized Qwen3.8-27B roles that only talk through files on disk to solve coding problems. On the first ten hard LiveCodeBench tasks the team averaged 9.8/10, surpassing Fable 5's 8.0 score.
Who it is for
Developers and AI researchers looking for a lightweight, locally-run multi-agent coding assistant can see how coordinated prompts boost performance.