DeepSeek-V4 at the same price as V3, on Huawei silicon, in preview

DeepSeek released preview versions of V4-Pro and V4-Flash on 24 April 2026. The release note is short, the pricing page is the part worth reading, and the most interesting sentence is about chips rather than parameters.¹

Start with the numbers, because they do most of the work. V4-Pro is a 1.6 trillion parameter mixture-of-experts model with 49 billion active parameters per forward pass and a 1M-token context window. V4-Flash is 284 billion total, 13 billion active, also 1M context. Both are MIT-licensed. V4-Pro's API pricing is $0.27 per million input tokens and $1.10 per million output tokens, with cache hits at $0.07. V4-Flash is $0.10 in, $0.40 out.¹

V4-Pro is a 1.6 trillion parameter mixture-of-experts model with 49 billion active parameters per forward pass and a 1M-token context window.

The pricing is the thing. DeepSeek-V3 launched in December 2025 at $0.27 input and $1.10 output. V4-Pro holds exactly the same price points at meaningfully larger scale, total parameters roughly double, active parameters up from ~37B to 49B, context window extended to 1M. DeepSeek is not expanding margin on a better model. It is absorbing scale into the same price line and letting MoE routing efficiency do the accounting.

This is the inference-economics frame doing what it does. The binding constraint at the frontier is no longer training cost; it is the per-token cost of serving inference at margin. DeepSeek has now run the same play three times, V2, V3, V4, each time holding or cutting the headline price while shipping a more capable model. Each cycle resets the cost floor for near-frontier open-weights inference. The frontier labs are pricing GPT-5.4 and Gemini 3.1-Pro at roughly an order of magnitude above this; whether that gap is justified by capability is the question every enterprise procurement team is now asked to answer in writing.

The capability gap is, per DeepSeek's own framing, three to six months behind GPT-5.4 and Gemini 3.1-Pro on standard benchmarks.¹ That is a self-reported figure and should be treated as such until independent evaluation lands, but the direction of travel is what matters. V3 at launch was further behind. The gap is compressing.

Now the chip sentence. From the release note: "DeepSeek-V4-Pro has full support for Huawei Ascend chips, enabling high-performance inference without Nvidia hardware."¹ This is the most geopolitically loaded line DeepSeek has shipped, and it is doing a lot of work in twelve words.

The US export-control architecture rests on a specific causal claim: restricting Nvidia hardware sales into China constrains Chinese frontier capability. If a 1.6T-parameter MoE runs at full performance on Ascend, the claim weakens. If it runs at degraded performance, the claim holds, with adjustment. The release note does not tell us which. There is no Ascend-versus-Nvidia throughput comparison, no latency figure, no specification of which Ascend generation. "Full support" is a phrase that can mean operationally equivalent or it executes. These are not the same thing and DeepSeek knows the difference.

I notice that DeepSeek has also not disclosed which chips were used to train V4. This is the second silence in the release and it matters more than the first. If V4 was trained on stockpiled or smuggled Nvidia H100s and merely runs on Ascend, the export-control story is mostly intact, Chinese labs can use Ascend for inference but still depend on Nvidia for training, which is the binding constraint. If V4 was trained on Ascend at scale, the architecture of the export-control regime is in trouble. The release note does not say. CNBC's reporting does not say. Until somebody benchmarks Ascend inference throughput against Nvidia and somebody else gets DeepSeek to disclose its training stack, "full support" is a claim, not a demonstration.²

What this is a case of: a Chinese frontier lab releasing a near-frontier open-weights model at the cost floor while making a chip-sovereignty claim that is structurally true if it is operationally true. The performativity frame applies. The claim itself moves markets, Nvidia equity, the export-control debate, enterprise China-strategy memos, regardless of whether the Ascend numbers, when they appear, fully support it. The geopolitical signal is doing work the technical disclosure has not yet done.

The MIT licence plus 1M context window is the third structural piece, and the one that matters most for Western API incumbents. V4-Flash at 13B active parameters is small enough to self-host on a manageable GPU footprint. Long-context RAG and document-processing workloads currently routed to OpenAI or Anthropic API endpoints have a substitution path that did not exist eighteen months ago: pull the weights, run them on owned or rented compute, pay zero marginal per-token cost. The pricing pressure on the API incumbents is not just DeepSeek's $0.27 input price. It is the fact that the price of self-hosting a near-frontier open-weights model is now bounded by GPU rental, not by anyone's API margin.

The contrarian read, which I want to give air to because it is correct on its own terms: this is a preview. DeepSeek has not given a GA date. Enterprise procurement does not move on preview releases. Prior DeepSeek previews have taken six to ten weeks to reach stable GA. The near-term competitive damage is to developer experimentation budgets and to the credibility of frontier-lab pricing, not to signed enterprise contracts. The benchmark gap is also self-reported. Independent evaluation usually narrows or widens such claims, rarely confirms them exactly.

Both things are true. The preview is not an enterprise displacement event this quarter. The trajectory is an enterprise displacement event over the next four quarters, if the cadence holds.

On cadence: V3 in December 2025, V3.2 shortly after, V4 preview in April 2026. Roughly four months from V3 to V4 preview. The intelligence-explosion frame asks whether successive releases compress the gap to the proprietary frontier. The arithmetic, taken at face value, is that a 3–6 month gap minus another four-month cycle takes the gap to zero by late 2026. I do not believe arithmetic this clean ever survives contact with reality. But the frame's prediction, that open-weights catches frontier within a manageable horizon, is not being falsified by this release. It is being mildly confirmed.

What to watch:

• Independent Ascend inference benchmarks against Nvidia H100/H200 on V4-Pro. Throughput and latency, not just functional execution.
• Disclosure (or leaked detail) on V4's training compute stack. This is the load-bearing question for the export-control debate, not the inference one.
• GA timeline and model card updates. The preview-to-GA window for prior DeepSeek releases.
• API incumbent pricing response. Watch for OpenAI or Anthropic moves on input pricing for long-context workloads in the next two quarters.
• Independent benchmark reproduction of the 3–6 month frontier gap claim.

The frame fits. The chip claim is the part that does not yet have evidence underneath it, and that is the part I would most like to see filed.

Footnotes

¶