DeepSeek V4 model release

The headline number — V4-Pro at roughly one-sixth the per-token cost of Claude Opus 4.7 and one-seventh the cost of GPT-5.5 — is easy to read as a market-share land grab funded by patient Chinese capital. The architecture tells a more interesting story. DeepSeek paired a token-wise compression layer with its DeepSeek Sparse Attention (DSA) mechanism so that a 1.6 trillion parameter MoE only activates 49 billion parameters per token, and at the 1-million-token context setting V4-Pro consumes just 27% of the single-token inference FLOPs and 10% of the KV cache footprint that V3.2 needed for the same workload. That is the actual reason the price sheet looks the way it does. DeepSeek is not selling intelligence at a loss — it has rebuilt the cost basis underneath it.

For builders, this matters more than the benchmark scores. A 1M-token context is largely theatrical at GPT-5.5 prices because nobody can afford to fill it; at $1.74 per million input tokens the long-context regime becomes the default rather than the special case. DeepSeek has also signaled further price cuts as Huawei's Ascend 950 supernode supply ramps in the second half of 2026, which means the V4 sticker is a ceiling rather than a floor. Frontier closed-source labs now have to decide whether to defend price or defend gross margin — and the open-weights MIT license under V4 means even the choice to defend price will not stop self-hosters from undercutting them in private deployments.

The framing in much of the U.S. coverage is that V4 was trained on Huawei chips, full stop. The reality, surfaced by Tsinghua's Liu Zhiyuan, is more measured: 'DeepSeek appears to have adapted only part of V4's training process for Chinese chips.' Some training stages still leaned on Nvidia silicon; what changed is that Huawei Ascend 950 hardware is now a credible part of the pipeline, and Huawei has publicly committed to full supernode deployment support around V4. SMIC shares jumping 10% on the news suggests the market read the signaling — not the engineering — as the headline event.

That distinction matters because the geopolitics flow from it. If V4 had been trained end-to-end on Ascend, U.S. export controls would be effectively obsolete on day one. What actually happened is closer to a credible threat: DeepSeek demonstrated that a flagship-scale model can incorporate Chinese chips into a frontier training run without collapsing in quality, and that subsequent generations can plausibly migrate further. Jensen Huang's reaction — 'the day that [DeepSeek comes out on Huawei] first, that is a horrible outcome for [the U.S.]' — reads less like a comment on this release and more like a forecast about the next one.

Independent benchmarking from Artificial Analysis places V4-Pro as the #2 open-weights reasoning model behind Kimi K2.6 and crowns it the top open-weight model on the GDPval-AA agentic benchmark with a score of 1554. Vals AI separately calls V4-Pro the '#1 open-weight model on our Vibe Code Benchmark, and it's not close.' By any normal reading, this is a state-of-the-art release. Then comes the asterisk: Artificial Analysis measured hallucination rates of 94% on V4-Pro and 96% on V4-Flash on probes designed to test whether a model knows what it does not know.

That is not a rounding error — it is a different kind of model than the benchmarks suggest. V4 appears to be optimized for performing well when it has the answer and confidently fabricating when it does not, which is exactly the failure mode that breaks agentic workflows in production. The practical read: V4 is excellent for tasks where outputs are immediately verifiable (code that compiles or fails, math that can be checked) and dangerous for tasks where the user assumes the model will say 'I don't know.' Anyone building retrieval pipelines on top of V4-Pro to exploit the cheap 1M context should pair it aggressively with grounding and citation checks rather than treating it as a closed-book oracle.

When DeepSeek-R1 dropped in January 2025, Nvidia lost roughly $600 billion of market capitalization in a single trading session and the entire U.S. AI capex thesis was briefly in question. The V4 release is structurally similar — a Chinese open-weights model leapfrogging the public price/performance frontier — but the market reaction has been an order of magnitude smaller. Nvidia fell only 1.41% intraday, SMIC rose 10%, and Chinese rivals MiniMax and Knowledge Atlas dropped 9% or more. Investors are repricing the competitive landscape inside Chinese AI, not the global compute thesis.

The difference is that the surprise has already been priced. Everyone now expects that DeepSeek and its peers will keep shipping cost-efficient frontier models on constrained compute; the existence of such a model is no longer the news. What V4 adds is the Huawei chip narrative and the IP-theft escalation from the Trump administration, both of which are political stories rather than market-structure stories. Michael Kratsios's framing — 'There is nothing innovative about systematically extracting and copying the innovations of American industry' — and the parallel rollout of new IP-theft accusations suggest the U.S. response will increasingly happen through trade and policy channels rather than through Nvidia's stock price. The shock is migrating from Wall Street to Washington.