DeepSeek V4 lands on Ascend

DeepSeek released V4 over the weekend: a 1.6-trillion-parameter mixture-of-experts model with a one-million-token context window, weights on Hugging Face under a permissive licence, and, the bit worth pausing on, first-class support for Huawei's Ascend processors out of the box.¹ The Pro variant places second on the standard world-knowledge suites, behind only Gemini-Pro-3.1, which is closed. Among open-weights models it is now the frontier. It reached the top of the Hugging Face trending list faster than any prior release.

The model card is, as is now customary at DeepSeek, terse to the point of comedy.² Three pages. A training-compute figure that I will return to. A benchmarks table. A licence. No safety section worth the name, a paragraph noting that "users are responsible for downstream deployment" and a gesture towards a refusal-tuning pass. The contrast with the Anthropic and OpenAI system cards, which now run to the low hundreds of pages, is becoming structural rather than stylistic.

The headline number everyone will quote is 1.6T parameters. The number that matters more is buried on page two: DeepSeek reports the model was trained "primarily on Ascend 910C clusters, with auxiliary H800 capacity for early-stage experimentation." Reported training cost: $84M.³ I have no way to verify the cost figure independently and the usual caveats apply, DeepSeek's numbers exclude the salaries, the failed runs, and the parent hedge fund's amortised infrastructure. But the Ascend claim is the load-bearing one, and it is testable: the inference stack ships with Ascend kernels, the quantisation recipes target Ascend memory hierarchies, and the model card explicitly thanks Huawei's CANN team. Either DeepSeek trained a frontier-adjacent model on domestic Chinese silicon, or they are pretending very specifically and in a way that Huawei would presumably object to.

Take the claim at face value for a moment. The export-control regime that the US has been tightening since October 2022 was designed around a thesis: deny advanced compute, slow the frontier, buy time. The thesis assumed that Chinese alternatives to Nvidia would lag by a generation or more, and that algorithmic efficiency could not close the gap inside the relevant window. V4 is not proof that the thesis is wrong, but it is the cleanest data point yet that it is wrong at the margin. DeepSeek is producing open-weights models that sit within touching distance of the closed US frontier, on hardware that the US cannot restrict, with training-compute figures that, even discounted heavily for optimistic accounting, suggest the algorithmic-progress multiplier is doing real work.

This is the intelligence explosion signals frame, and I'd note where the frame fits and where it doesn't. It fits in the sense that compute-efficiency gains are clearly compounding: V4 reports roughly 3.5x the effective training throughput of V3 on comparable hardware, which is the kind of multiplier the frame predicts. It fits in the sense that the export-control lever is visibly losing torque. It does not fit cleanly as a "race" data point, because DeepSeek is not racing anyone in the way the frame imagines, they are releasing weights, which is the opposite of the closed, secured, government-aligned posture that the race framing assumes from a national champion. China's frontier lab is, structurally, behaving more like Meta than like OpenAI.

The timing, then. V4 dropped the same week the White House published an accusation that Chinese labs are systematically exfiltrating IP and weights from US frontier developers.⁴ I am not going to litigate that here, the public evidence is thin and the classified evidence is, by definition, classified. But the market structure point is that the accusation lands awkwardly when the accused has just released a model whose architecture is visibly its own (the MoE routing scheme is novel, the long-context attention variant is described in the accompanying paper, and the Ascend-native training stack is plainly not stolen from anyone running H100s). You can believe both that some IP theft is occurring and that V4 specifically is not the artefact of it. The administration's statement implicitly conflates these.

For the AI safety as market position frame: this is the cleanest divergence yet. Anthropic's pitch to enterprise and to government is that safety posture is a feature you pay for. OpenAI is converging on the same pitch, more reluctantly. DeepSeek's V4 release demonstrates that there is a viable parallel market in which safety posture is not a feature, weights are public, and the model is good enough. The question for the closed frontier labs is no longer whether the open-weights frontier exists, it does, but how much of the enterprise market they can hold on the safety-and-compliance proposition once the capability gap, currently maybe nine months, compresses further. I'd watch the regulated verticals. Banks and hospitals are not going to deploy V4 directly. But the systems integrators who wrap it, fine-tune it, and sell it as a sovereign-deployable alternative are going to have a very interesting 2026.

The inference-economics angle is more straightforward and worth stating cleanly. V4 weights are free. Inference is not. The Ascend stack means inference can be run domestically in China at costs that, on the numbers Huawei discloses, are competitive with H100-class deployments. The pricing pressure this creates on US inference providers is real but second-order, most enterprise customers in the West cannot or will not deploy on Ascend. The first-order pressure is on the open-weights inference market in the West, where V4 will be served by the usual suspects (Together, Fireworks, Groq, the hyperscalers' own endpoints) and will compete directly with served Llama and served Mistral on price. I'd expect the per-token price for frontier-adjacent open-weights inference to step down materially within the quarter.

What this is a case of: the increasingly clear pattern in which the open-weights frontier is being set by labs that are structurally outside the American safety-and-alignment consensus, on hardware that is structurally outside the American export-control perimeter, at training costs that, whatever the precise figure, are an order of magnitude below what the closed frontier spends. The frames that assume a single frontier with a clear leader are aging quickly. There are now two frontiers, with different economics, different release norms, and different customer bases, and the gap between them is a capability gap measured in months rather than the structural moat the closed labs were assuming a year ago.

What to watch: the Ascend claim getting independently reproduced or debunked by someone running the training recipe; the per-token pricing on V4 endpoints over the next four weeks; whether any US-regulated enterprise actually deploys V4 in production, and under what compliance wrapper; and the next Anthropic or OpenAI system card, to see whether the safety-posture-as-product pitch sharpens or softens in response.

Footnotes

¶