DeepSeek V4 and V4-Pro: what changed this year
DeepSeek's 2026 releases from its own changelog, what V4 and V4-Pro are, what the MIT-licensed weights mean for a business, and where your data may go.
DeepSeek shipped a run of model releases in 2026, and the names change quickly: V4, V4-Pro, V4-Flash, V4.1-Flash, each with dated builds. For a business, the names matter less than three questions. What can the models do, what do they cost to call, and where does your data go when you use them?
Here is the year as DeepSeek’s own changelog records it, what V4 and V4-Pro actually are, and how we would weigh them for a small business. Benchmark figures below are DeepSeek’s own, not independent tests.
The 2026 timeline
DeepSeek’s API changelog lists the releases below. All dates are in 2026.
| Date | Release | What changed |
|---|---|---|
| 24 April | V4 launch: deepseek-v4-pro and deepseek-v4-flash | Two models in the new V4 family, both with a one million token context window |
| 31 July | deepseek-v4-flash public beta | Same architecture as the earlier V4-Flash preview, post-trained again |
| 13 August | V4-Pro general release (build 0813) | The official V4-Pro, replacing the preview |
| 16 August | Peak and off-peak pricing starts | Off-peak rates are half of peak rates |
| 21 August | deepseek-v4-flash-vision-exp | Experimental model that accepts images through the API |
| 10 September | V4.1-Flash release (deepseek-flash) | The smallest model in the family, with native image understanding |
| 17 October | V4.1-Flash-1017 update | Same API name and pricing according to the changelog, and the Search API opens |
The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the models behind them have been retired, and their requests now go to the 1017 release. The original deepseek-chat and deepseek-reasoner names were scheduled to be discontinued on 24 July. If your software calls any of these names, check which model it actually reaches before you rely on its answers.
What V4 and V4-Pro are
DeepSeek-V4 launched on 24 April 2026 as two models, V4-Pro and V4-Flash. Both use a mixture-of-experts design: the model is built from many specialist parts, and only a few of them work on each token of text. That lets a very large model do less work per word than its total size suggests, which is why these models are quoted with two sizes.
According to the April model cards, V4-Pro has 1.6 trillion parameters in total, with 49 billion active for each token. V4-Flash has 284 billion in total, with 13 billion active. The card for the August general release of V4-Pro lists 1.7 trillion, so DeepSeek’s own pages give two totals for the same model.
Both V4 models accept up to one million tokens of context. The context window is how much text the model can consider in one request, including any documents you paste in. A large window lets you send a whole contract or a year of meeting notes at once, but every token you send is input you pay for.
V4-Pro is the larger model, aimed at harder reasoning and coding work. V4-Flash is the smaller and cheaper option. V4-Pro reads text only; image input arrived on the Flash side, first as the August experiment and then in V4.1-Flash.
Open weights under the MIT License
DeepSeek publishes the model weights on Hugging Face, the public site where AI model files are shared. The cards for V4.1-Flash and the August V4-Pro build both state: “This repository and the model weights are licensed under the MIT License.” MIT is one of the most permissive licences in common use, and it allows commercial use. Our earlier post on open weight and open source explains what open weights do and do not give you.
Downloading the weights is not the same as running them. The model cards give serving examples for vLLM and SGLang, two common model-serving tools, but they do not state hardware requirements. At 1.6 trillion parameters, self-hosting V4-Pro is a specialist project. For most small businesses, the hosted API is the practical route, which brings us to cost and data.
Price and performance
DeepSeek’s case for V4 is strong results at a low price. On the April V4-Pro card, DeepSeek reports 80.6 on SWE Verified, a test of fixing real software bugs, and 90.1 on GPQA Diamond, a set of hard science questions. These are DeepSeek’s figures, run at its highest reasoning setting. A later post in this series looks at the agent results that came with the August release.
On price, DeepSeek’s pricing page lists the current Flash model, deepseek-flash, at $0.15 per million input tokens and $0.60 per million output tokens off-peak, and $0.30 and $1.20 at peak. V4-Pro, deepseek-v4-pro, lists $0.66 and $1.98 off-peak, and double that at peak. These are the prices at the time of writing (late October 2026). The page shows a dollar sign without a currency code, so confirm the currency there before you budget.
The unit price is only part of the cost. A cheap model that gets a task wrong costs more in checking time. Take ten or so real examples from a task your team already knows well, run them through two or three models, and compare the answers with what your people would have produced. When a model passes that test, the next step is connecting it to the systems you already run, which is the kind of software development work where the data rules need to be agreed first.
Where your data goes
Price is the easier question. Data takes more care. DeepSeek’s consumer privacy policy, last updated 10 February 2026, says personal data may be stored on a server outside the country where the user lives, and that DeepSeek collects, processes and stores personal data in the People’s Republic of China. It lists training its models as one purpose of collection, and gives users the right to opt out of that.
Two points keep this in proportion. First, the policy predates V4 and covers the consumer service; it does not say plainly how API traffic is handled. Second, the open weights change the picture: if you run the model yourself, or with a hosting provider you choose, your requests never go to DeepSeek at all.
So the practical rule is simple. Public content and made-up test data are fine for a trial on the hosted API. Before any personal data or confidential records go through it, ask DeepSeek which terms apply to API requests, how long it keeps them, and whether it offers a data processing agreement, and get the answers in writing. A later post in this series compares the hosted API with running the weights yourself.
Sources
Start with the people doing the work.
We build software, AI chatbots, and AI agents around the systems you already run, with people keeping the judgment.