DeepSeek’s smaller model just outperformed its own flagship

DeepSeek just did something a little counterintuitive: it made its smaller model beat its bigger one. The company pushed out DeepSeek-V4-Flash-0731 through a public beta API last week, and dropped the open weights on Hugging Face under an MIT license the same day. No architecture changes. Just additional post-training, according to DeepSeek, and that alone was enough to leapfrog the earlier V4-Pro preview on several agent benchmarks.

The numbers make the gap concrete. V4-Flash runs on 284 billion total parameters with 13 billion activated per token, versus V4-Pro's 1.6 trillion total and 49 billion activated. That's not a small difference in inference cost — it's nearly a 4x gap in activated parameters, which matters a lot if you're running agents at scale and paying by the token. DeepSeek reported 82.7 on Terminal-Bench 2.1, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified for the new Flash model. Artificial Analysis, running its own independent check, landed on 79% for Terminal-Bench 2.1 — close, but not identical, which is a useful reminder that vendor-reported benchmarks deserve a second look before anyone bets infrastructure budget on them.　

Still, even with that discrepancy, the broader signal holds: you don't always need a bigger model to get better results. Post-training tweaks apparently did real work here, and if DeepSeek's other internal numbers hold up under outside scrutiny, that's one more data point in a growing pile suggesting labs are squeezing more out of existing weights rather than just scaling up.

The license move matters just as much as the benchmark story. MIT-licensed weights mean companies can self-host V4-Flash instead of being tied to DeepSeek's hosted API, which is a meaningfully different proposition than most closed-model vendors offer. Pair that with Responses API support and published instructions for plugging into Codex-based workflows, and DeepSeek has made switching costs almost trivially low for teams already building on OpenAI-style tooling. They don't have to rearchitect anything to test it.

DeepSeek also mentions inference-side work — speculative decoding through its DSpark framework — aimed at making self-hosted serving cheaper and faster. Combine that with the open weights and the familiar API shape, and what's emerging is less a single-model release and more a statement about where the competition is heading: not purely on raw capability, but on how flexible and cheap it is to actually run the thing in production.