DeepSeek ships V4-Pro to general availability

DeepSeek moved V4-Pro to general availability on August 13, 2026. The GA build, snapshot DeepSeek-V4-Pro-0813, now serves the app, the website, and the API under the existing name deepseek-v4-pro, so API callers were switched over without changing anything. The weights are on Hugging Face under an MIT license.

V4-Pro had been in preview since April 24, when DeepSeek released it alongside a smaller sibling, V4-Flash, as open weights: 1.6 trillion total parameters with 49 billion active for Pro, 284 billion and 13 billion for Flash, both with a 1 million token context window. Hugging Face lists the 0813 checkpoint at 1.7 trillion parameters. According to the model card, the GA build keeps the preview's structure and adds DSpark, a speculative decoding module: a small companion drafts tokens ahead and the main model checks them, which speeds up generation.

How much better is it than the preview?

DeepSeek's pitch is agents, meaning software that calls tools, writes code, and finishes a multi-step job without a person approving each step. Every number here is the company's own, produced in its own agent framework, DeepSeek Harness, in minimal mode at the max thinking setting. No independent lab has published matching results for 0813 yet.

• Terminal-Bench 2.1, which tests whether a model can operate a real command line: 87.9, up from the preview's 72.1 according to the model card's table.
• DeepSWE, which scores real software engineering work: 62.7, up from 12.8. That jump is the headline of the release.
• Humanity's Last Exam, a deliberately hard general knowledge test: 42.7 without tools, 60.0 with them.

The published table also lists Cybergym at 83.3, NL2Repo at 61.5, and Toolathlon-Verified at 74.1, all in the same agent vein. DeepSeek calls the model broadly competitive with the strongest proprietary models, which is exactly the kind of claim that needs outside testing.

There are practical additions too. Thinking effort, the amount of internal reasoning the model does before answering, can now be set to low, high, or max on both V4-Pro and V4-Flash. And the API natively accepts the OpenAI Responses format and has been adapted to work with Codex, OpenAI's coding agent, a fairly direct invitation to point OpenAI-shaped tooling at DeepSeek's endpoint instead.

What does it cost?

Pricing changes on August 16 at 16:00 UTC, when DeepSeek starts billing by time of day: off-peak hours cost half of peak hours. The company has not stated the per-token rates. For agent workloads, which burn tokens by the million, a batch job that can wait for off-peak hours pays half as much as a session that cannot.

Can you run it yourself?

In principle, yes. The 0813 weights are on Hugging Face under MIT, with support for the vLLM and SGLang serving stacks and quantization options for local deployment. A frontier-class flagship you can call on a public API and also download outright is still uncommon. But at 1.7 trillion parameters this is data center hardware, not a workstation.

What is still unknown

Independent benchmark results for 0813 do not exist yet, so the scores above remain company claims. The dollar rates under the new peak and off-peak scheme are unpublished. And DeepSeek has not said whether a later snapshot will replace 0813 under the same deepseek-v4-pro name, as its earlier model families have done.