DeepSeek's Flash outpaced its own flagship. The upgrade was post-training, not parameters.

DeepSeek released V4-Flash-0731 last week, maintaining the same 284B parameter architecture and 13B activated parameters per token as the preview version. This upgrade did not introduce any architectural changes or a larger model; instead, the gains came from post-training improvements to the agent capabilities. The model's performance surpassed V4-Pro-Preview on several agent benchmarks, making this release noteworthy for its method rather than the model itself.

The activated-parameter gap between Flash and Pro is significant, as inference cost scales with activated parameters rather than total parameters. Flash is running at roughly a quarter of the activation cost of Pro while also outperforming Pro on agent tasks. Benchmarks include 82.7 on Terminal-Bench 2.1, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified, with Artificial Analysis reporting a Terminal-Bench score of 79%, highlighting the importance of independent verification.

The MIT license allows full self-hosting rights, no API dependency, OpenAI-style API compatibility, Codex workflow integration, and DSpark speculative decoding, which claims an 85% inference speed improvement for self-hosted deployments. Running agents on frontier models, especially with Flash-0731's agent-specific post-training, may offer lower costs and better performance, particularly for tool-call heavy workflows.

Switching to OpenAI-compatible APIs is low-risk, and self-hosting with the MIT license and DSpark is a credible production stack. However, skepticism is warranted, and further independent replication of the benchmark claims is advised.