DeepSeek V4: the benchmark is the distraction. The chip pivot isn't.

Let's get the easy part out of the way. On 24 April 2026 DeepSeek released two open-weight models under an MIT licence: V4-Pro (1.6 trillion total parameters, 49 billion active) and V4-Flash (284 billion total, 13 billion active). Both support 1 million-token contexts. V4-Pro lists at $1.74/M input (it launched on a 75% promo at $0.435/M); V4-Flash at $0.14. Weights are on Hugging Face, self-hostable, commercial-use allowed. Artificial Analysis put V4-Pro at #2 on the open-weight reasoning index, behind Kimi K2.6. Impressive? Yes. 'Best model in the world'? No — the CFR pegs the U.S. as still roughly seven months ahead. Those are the receipts.

Here is the comparison the leaderboard crowd is missing:

(V4-Pro launched on a 75% promo at $0.435/M input.) The context jump from 128K to 1M tokens between V3.2 and V4-Pro is real and matters for enterprise use cases. The inference efficiency story — 27% of V3.2's FLOPs — is also real. Neither of these is the headline in most V4 coverage, and both will outlast the benchmark comparison by 18 months.

Every previous DeepSeek model ran on Nvidia hardware. V4 does not. Per MIT Technology Review, DeepSeek reportedly gave early access only to Chinese chipmakers; Huawei announced same-day Ascend 950 support — not a coincidence. What this tells you is that DeepSeek is engineering its future model development around a chip ecosystem the U.S. cannot cut off via export controls. Whether Ascend 950 is as efficient as H100 for inference is a separate technical question that will be answered in the months that follow. The strategic intent is not ambiguous.

The CFR analysis — published five days after launch — is worth reading in full. Its core argument is that the right frame for V4 is not 'does it beat GPT-5?' but 'will it be the default model for the next generation of AI applications built outside the U.S. and Europe?' At $1.74/M input with self-hosting rights, the answer could easily be yes for a significant portion of global developers. That is the strategic consequence, and it is not resolved by whether V4-Pro ranks #1 or #2 on an intelligence index.

File V4 correctly: a genuine open-weight leap, not the state of the art, priced to win the global adoption race, and built for the first time on hardware the U.S. can't sanction. DeepSeek claims V4-Pro tops some coding tests outright — its own report lists a LiveCodeBench score of 93.5 versus Claude Opus 4.6's 88.8, though that's a self-reported, not-yet-independently-confirmed number — yet on the independently administered composite Intelligence Index V4-Pro still sits second to Kimi K2.6, which is exactly the point: this is an excellent open model, not a frontier-beater, and its leverage is cost and openness rather than a leaderboard crown. The benchmark column is interesting. The chip column is load-bearing.