DeepSeek V4-Pro leaves preview with MIT weights and ranks third in independent benchmarks

August 12, 2026. DeepSeek moved its flagship V4-Pro from preview to general availability, under build 0813. August 13, 2026. The model weights landed on Hugging Face under an MIT license. Artificial Analysis published its independent verdict soon after: a score of 53, third in the global ranking. The Chinese flagship is now a real open-weight model, not a promise.

The episode closes a sequence we have been tracking: in July, the V4-Flash already beat the V4-Pro preview through re-post-training alone, and the peak/off-peak rate card had changed the sector’s economics. The GA release of V4-Pro finishes the trilogy: the big model is now here, open, and measured.

What GA actually changes

The April 2026 preview was a teaser. The August 13 GA is a product. V4-Pro-0813 keeps the announced architecture: 1.6 trillion total parameters, 49 billion active per token, a Mixture-of-Experts design with Compressed Sparse Attention and Heavily Compressed Attention, pre-trained on more than 32 trillion tokens. The advertised context is 1 million tokens with up to 384,000 of output, and three reasoning modes — non-thinking, high reasoning, and max reasoning.

The concrete difference is distribution. The weights sit on Hugging Face under MIT, alongside a DSpark variant that ships a native speculative decoding module for vLLM and SGLang — no separate draft model needed to speed up inference. For a research lab, or a team that wants to audit weights rather than consume an API, this is the first time a model of this size, at the top of the rankings, downloads without a restrictive license. The built-in speculative decoding is also a clever engineering choice: it removes a piece of infrastructure to maintain, at the cost of a slightly heavier checkpoint.

Under the hood: the MoE economics

Why activate only 49 billion parameters out of 1.6 trillion? The answer is inference. A dense model reads every weight on every token; an MoE reads only the experts relevant to the current token, plus a router that chooses them. DeepSeek combines two techniques to compress attention — Compressed Sparse Attention and Heavily Compressed Attention — which shrink the context-cache memory needed on long sequences.

The gain shows up in dollars per million tokens, not abstract percentages: at comparable capability, reading 49 billion weights instead of 1.6 trillion cuts the compute bill by an order of magnitude. That is what lets V4-Pro list an output price of $3.96 while a dense frontier model of the same caliber charges $30. The hidden cost is memory: the 1.6 trillion weights still have to be loaded somewhere, which pushes self-hosting toward multi-node clusters.

The independent verdict, more interesting than the model card

DeepSeek’s model card shows aggressive scores: 80.6% on SWE-bench Verified, 90.1% on GPQA Diamond, 93.5% on LiveCodeBench, 87.5% on MMLU-Pro, 67.9% on Terminal Bench 2.0, and a Codeforces rating of 3,206. Artificial Analysis added a third-party measurement.

The result: 53 on its Intelligence Index, the third place globally. That is unambiguously frontier territory. But the gap to the small V4-Flash is three points — from a model activating five times more parameters per token. The honest read fits in one sentence: V4-Pro is genuinely a top-three model, at a fraction of frontier prices, but the card’s boldest claims still need validating on your own workload. Cybersecurity stands out as the model’s strength in early third-party coverage — a notable signal in a market crowded with models tuned for coding benchmarks.

Open weights are not free

Calling V4-Pro “open” is accurate in license terms — the weights are MIT. It is not accurate in resource terms: a 1.6-trillion-parameter checkpoint is multi-node territory, out of reach of a homelab. The MoE design follows a constant rule: fewer active parameters save compute, not memory — the full weights still have to fit in VRAM.

The rough math: 1.6 trillion parameters at bf16 means on the order of 3.2 TB of weights before any KV cache or overhead — a multi-node GPU cluster, not a workstation. V4-Flash, at 304 billion, lands closer to 600 GB, still beyond a single high-end card but within reach of a small server.

In practice, most teams will consume V4-Pro through the API or a host, and reserve self-hosting for V4-Flash, whose 304 billion parameters are themselves multi-GPU. The immediate win is therefore not hardware independence — it is reversibility: being able to leave the API without losing access to the model, and to audit what you run. In a context where vendor dependence is becoming a governance risk, that reversibility is sometimes worth more than the token savings.

Where V4-Pro sits against the competition

The positioning reads less from the model card than from the price-to-capability gap against other flagships. V4-Pro lists an output price of $3.96 per million tokens at peak ($1.98 off-peak). Against that, Qwen 3.8-Max runs $6, Kimi K3 $15, and GPT-5.6 Sol $30 — for capability classes in the same ballpark. DeepSeek’s reasoning is plain: be top-three at a tenth of the frontier price, and let the price do half the pitch.

That strategy has a structural consequence: it compresses margins across the entire “model” layer. For a CTO arbitrating between vendors, the question is no longer “who is best” but “at what quality gap does the premium become indefensible”. The answer, more and more often, is: never. That is the pressure Chinese open weights put on the rest of the market, beyond any technical debate.

What it means for the open-weight ecosystem

The GA of V4-Pro reinforces a broader shift. Since the start of 2026, Chinese open-weight models — DeepSeek, Qwen, Kimi — have closed the gap with closed American frontiers, to the point where the best open model is now a category enterprises track as a market of its own. Publishing V4-Pro’s MIT weights is not philanthropy: it is a distribution strategy. The more the model is copied, hosted, and audited, the more the surrounding tooling ecosystem builds around DeepSeek, and the less customers depend on a single vendor.

For a CISO or a CTO, the practical consequence is this: open weight is no longer a risky bet on an outsider, it is a sovereignty option — the ability to run the model in your own infrastructure or at a host of your choice, and to audit what you execute. The trade-off remains inference: sovereignty is paid for in GPU.

Verdict

If you are evaluating a frontier model for agentic or coding tasks, add V4-Pro-0813 to your matrix: it is a top-three measured by an independent source, and its cybersecurity strength deserves a dedicated test. If you want to audit or self-host, download the MIT weights — but size a cluster, not a server. If price is the driver, benchmark on your own load: the three-point gap to V4-Flash, at lower cost, may be enough for many uses. The model card describes the ceiling; your tests describe the floor.

References

• DeepSeek V4: Release Date, Specs, and How to Access It (2026) — Yotta Labs, August 27, 2026
• DeepSeek officially launches V4-Pro AI model in August 2026 — Reuters via Yahoo Tech
• DeepSeek V4 Pro: Specs, Flash Comparison & Pricing — deepseek-v4.io
• DeepSeek V4 Guide: Pro & Flash, GA + Pricing (Aug 2026) — Codersera