NVIDIA Took DeepSeek's V4 Pro 14 Days After Launch

The Model Card That Belongs to Someone Else

There is a page on Hugging Face that tells most of the story: nvidia/DeepSeek-V4-Pro-0813-NVFP4, uploaded August 27, 2026, MIT license, quantized with NVIDIA's own Model Optimizer tool. The model card is candid about what it is. "This model is not owned or developed by NVIDIA," it reads — the model was "developed and built to a third-party's requirements."

In other words: NVIDIA's organization is distributing a flagship model from a Chinese company NVIDIA does not own, did not build, and did not train. Fourteen days after its official launch. Six likes so far. And yet this page may be the most revealing artifact of the entire release cycle.

What V4 Pro-0813 Actually Is

On August 13, 2026, DeepSeek formally ended a preview period that had been running since April. V4 Pro-0813 is the official edition of the flagship: a Mixture-of-Experts model with 1.65 trillion total parameters, of which 49 billion activate per token. The architecture combines hybrid attention — Compressed Sparse Attention and Heavily Compressed Attention — with what the company calls Manifold-Constrained Hyper-Connections. The context window runs to one million tokens.

Two details in the checkpoint reveal what DeepSeek is betting on. The DSpark speculative decoding module ships inside the weight file itself, not as a separate software layer — faster inference is baked into the deliverable. And the model natively supports the OpenAI Responses API, with reasoning effort levels of low, high, and max. This is a model card written for enterprise agents, not for benchmark chasing.

The Mixed Reception

The "DeepSeek strikes again" framing does not hold this time. Reuters described the launch as an attempt to regain ground against fast-moving domestic rivals while the company simultaneously expands hiring and compute. That is a position of pursuit, not dominance.

The South China Morning Post made the numbers grimmer than the press release: developers were underwhelmed by the model's overall benchmark performance, even as it excels in cybersecurity. That is a pattern worth noting — the update landed unevenly across domains rather than lifting everything at once. A flagship relaunch that gleams narrowly and disappoints broadly is a different story from another shockwave moment.

The Pricing: From Price Pressure to Cash Flow

The clearest strategic signal is the price. Reuters reports that V4 Pro launched at rates up to 14 times higher than V4 Flash. That is a spread, not a single price — the tiers separate the front-line product from the volume product — but the direction is unambiguous: DeepSeek is trying to convert prestige into willingness to pay.

At the same time, eWeek stresses that pricing remains well below leading competitors. That is the trick in the 14x figure: V4 Pro is no longer a giveaway, but it is not premium-priced against the front-line American models either. DeepSeek has built a hierarchy — V4 Flash as the volume catcher, V4 Pro as the margin catcher for the agentic workloads enterprises actually pay for.

This connects to the other launch component: an open-source evaluation tool for agentic tasks, and benchmark gains precisely there. Differentiation is moving from raw intelligence to the deployment layer — API compatibility, speculative decoding, agent queues. That is where the customers sit, and where prices can be defended.

NVIDIA's Quiet Hand

Now to the quantized copy. NVIDIA uploaded the NVFP4 version on August 27 — two weeks after the official launch. Technically, it is straightforward: Model Optimizer compresses the weights to 4-bit floating point, so a 1.65T-parameter model with only 49B active parameters can run more cheaply on NVIDIA silicon.

The mechanism is worth pausing on. DeepSeek's open weights, released under an MIT license, mean anyone can take the checkpoint. NVIDIA's version lowers inference cost on exactly the machines NVIDIA sells. Open frontier models, in other words, do not become an alternative to NVIDIA hardware — they become sales arguments for it. Every time a lab releases its weights openly, the ecosystem produces free material that makes NVIDIA's platform more attractive.

That is the distribution story in V4 Pro. The capability landed fast. The reception was mixed. The pricing shifted toward monetization. And while DeepSeek tried to profit from its flagship, the infrastructure vendor pulled the checkpoint into its own pipeline — the MIT license allowed it, no more and no less.

The Counterweights

Honesty requires three caveats. First: the mixed benchmark performance suggests the model does not yet justify premium pricing on capability alone. SCMP's finding of underwhelmed developers is a real counterweight to the narrative of another frontier shift.

Second: the domestic rivals Reuters points to are moving fast, and the 14x price spread gives them room to attack from below.

Third: NVIDIA's upload has six likes on Hugging Face. That is not adoption; it is an infrastructure move — a signal of direction, not a traffic number. Reading the quantized copy as proof of market power would repeat the same mistake as reading benchmark scores as product features.

Who Owns the Value

Back to the model card. A model built to a third party's requirements, released under a license that lets anyone take it. DeepSeek got its launch, its pricing, and its place in the conversation. NVIDIA got the checkpoint — and a new, free proof that the heaviest open models pay off best when run on NVIDIA's terms.

The question the launch leaves behind is not whether V4 Pro beats its rivals on benchmarks. It is who owns the value when frontier weights are free: the lab that trained them, or the toolmaker who makes them cheap to run. The fourteen days between the two uploads offer a hint.