DeepSeek's V4-Flash-0731 Jumped From 7% to 54% on DeepSweep Through Post-Training Alone — Same Architecture, Same Parameter Count

DeepSeek·low signalXBlueskyLinkedInCopy link

DeepSeek shipped production V4-Flash-0731, a 284B MoE with 13B active parameters and a 1M-token context, keeping the exact structure and size of V4-Flash-Preview and redoing only post-training. The result: agentic coding score on DeepSweep went from 7% to 54%, and DeepSeek reports the release beating its own larger V4-Pro-Preview across nine agent benchmarks, plus beating GLM 5.2 on nearly every published benchmark despite GLM running roughly three times the parameter count. The lesson for builders is that agentic capability is currently post-training-limited, not scale-limited — the same weights class went from unusable to competitive without a new pretrain.