DeepSeek released new open weights for DeepSeek-V4-Flash on July 31, and the change is entirely in the post-training. The architecture, the million-token context window and the expert layout are identical to the preview version, yet the model's score on DeepSWE - a test that asks an agent to resolve real issues in real software repositories - went from 7.3 to 54.4 out of 100. The weights are on Hugging Face under an MIT licence, and the existing API name now points at the new checkpoint with no code change required.
Key facts
- DeepSeek-V4-Flash-0731 scores 54.4 on DeepSWE versus 7.3 for the preview checkpoint, on DeepSeek's own harness.
- Released July 31, 2026, as MIT-licensed open weights that supersede the preview version.
- The API identifier
deepseek-v4-flashnow resolves to 0731; pricing is $0.14 per million fresh input tokens, $0.28 per million output tokens, and $0.0028 per million cached input tokens. - Primary source: the official DeepSeek-V4-Flash-0731 model card and the DeepSeek API pricing page.
The number that makes people stop scrolling is not 54.4. It is the gap between 7.3 and 54.4 with nothing structural in between.
To understand why that is strange, it helps to know how a model like this is built. There are two broad phases. Pretraining is where the model reads an enormous amount of text and learns the shape of language and knowledge; it is where nearly all of the compute and money goes, and it is what determines the parameter count everybody quotes. Post-training is the shorter, cheaper phase afterwards, where the model is taught how to behave: follow instructions, use tools, stop when finished, admit uncertainty. Conventional wisdom treats pretraining as the part that sets a ceiling and post-training as polish.
DeepSeek's release argues that for agent work, the polish was most of the ceiling. Its model card states plainly that 0731 "outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count" - that is, the cheaper, smaller-compute sibling now beats the flagship on the agent tests DeepSeek chose to publish. On the same table, Terminal Bench 2.1 moves from 61.8 to 82.7 and a cyber-reasoning test called Cybergym moves from 38.7 to 76.7.
The mechanism underneath is a mixture-of-experts design. The model holds 284 billion parameters in total but routes each token through only about 13 billion of them - the equivalent of a very large reference library where any single question sends you to a handful of shelves rather than the whole building. That is why the per-token price can be low while the stored model stays enormous. It is also why "13 billion active" is a misleading shorthand for anyone hoping to run it at home: the shelves you do not open still have to exist. Unsloth's community quantizations start around 83GB at aggressive low precision and reach roughly 162GB at 8-bit.
The pricing is where the release lands hardest on working developers. Fresh input costs $0.14 per million tokens; output costs $0.28; and input that DeepSeek has already seen and cached costs $0.0028, a fiftyfold discount on the same tokens. For an agent that re-sends a large, stable system prompt and codebase on every turn, that cache rate is the whole economic story. DeepSeek also notes on the same page that it will "soon adopt a peak/off-peak pricing policy" doubling prices during Beijing business hours, with the start date still unannounced - so today's arithmetic has a scheduled expiry.
Independently, Artificial Analysis scored the new checkpoint at 50 on its Intelligence Index, ten points above the previous Flash, which puts it within a point or two of GLM-5.2 and GPT-5.6 Luna while remaining well behind Kimi K3. That is the honest placement: near the frontier's lower shelf, at a fraction of the frontier's price.
The caveat is the one DeepSeek itself documents. Every code-agent number in that table was produced with something called the DeepSeek Harness, in minimal mode, at maximum reasoning effort - and the card says the harness is "to be released". Two of the listed tests, DSBench-FullStack and DSBench-Hard, are internal sets nobody else can run. The public DeepSWE leaderboard puts every model through one shared scaffold precisely to prevent this problem, and it has not scored Flash 0731. Until it does, 54.4 is a strong vendor claim about a model plus its own unreleased toolkit, not a rank. The jump is real and the price is verified; the leaderboard position is not yet a thing that exists.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)