DeepSeek moved V4-Flash out of preview on July 31, and the changelog entry for the deepseek-v4-flash endpoint carried a single line worth reading twice: the new V4-Flash-0731 build "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained." Same 284 billion total parameters, same 13 billion active per token, same 1M-token context, same MIT license, same $0.14-per-million-input and $0.28-per-million-output pricing. What changed is what happened after pre-training. And it changed a lot.

On the nine agent and coding benchmarks DeepSeek published, the re-post-trained Flash beats V4-Pro-Preview, the company's own flagship. Terminal-Bench 2.1 climbs from 61.8 on the preview Flash to 82.7 on 0731, versus 72.1 for V4-Pro-Preview. DeepSWE goes from 7.3 to 54.4, a 645% jump DeepSeek attributes to reinforcement learning with verifiable rewards against executed code.

The caveat is real. The DeepSWE number is vendor-reported, run through DeepSeek's own Harness framework in "minimal mode" at max effort, top-p 0.95, temperature 1.0. DeepSeek says the harness will be released "soon"; until it's, nobody else can reproduce the run. An independent DeepSWE evaluation by yage.ai clocked V4-Pro at 8% pass@1, against 70% for GPT-5.5 and 54% for Claude Opus 4.7, a reminder that vendor harnesses and third-party harnesses aren't the same instrument.

Artificial Analysis, which does run its own harness, has V4-Flash-0731 at 50 on its Intelligence Index, 10 points above the previous Flash and one point behind GPT-5.6 Luna (max) at 51. It notes the model "shares identical architecture and pricing with the earlier DeepSeek V4 Flash." Cost per Task on DeepSeek's first-party API lands roughly 60% below Luna, even after OpenAI's recent 80% price cut, helped along by a cache-hit discount of about 98% versus the industry's typical 90%.

The Omniscience numbers tell the cleanest story about what post-training can and can't move. AA-Omniscience rises 7 points to −16, hallucination rate falls 11 points to 84%, and accuracy sits unchanged at 37%, which Artificial Analysis reads as consistent with an unchanged parameter count. You can teach a fixed-size model to stop guessing. You can't teach it to know more.

That's the structural point buried in a routine changelog. The 2023–2024 discourse assumed the frontier moved by scaling parameters and pre-training compute; the 2025 discourse assumed it moved by adding reasoning tokens at inference. DeepSeek's July 31 checkpoint suggests a third axis, and a cheaper one: hold the weights, redo the post-training, ship the same endpoint. The budget model quietly outscores the flagship, and the flagship's name doesn't appear in the release notes at all.

DeepSeek V4-Flash-0731: A Re-Post-Trained Budget Model That Beats Its Own Flagship

Sources