DeepSeek's V4-Flash-0731 jumped from 7% to 54% on DeepSweep through post-training alone.

Same architecture, same parameter count: a 284B MoE with 13B active and 1M-token context, keeping the exact structure and size of V4-Flash-Preview with only post-training redone (DeepSeek). Agentic coding score went 7% → 54%. DeepSeek reports the release beating its own larger V4-Pro-Preview across nine agent benchmarks, and beating GLM 5.2 on nearly every published benchmark despite GLM running roughly three times the parameters. The read for builders: agentic capability is currently post-training-limited, not scale-limited. Same weights class went from unusable to competitive without a new pretrain. If you're waiting for a bigger model to fix your agent, you may be waiting for the wrong thing.