DeepSeek V4 is the model family that put DeepSeek back in the headlines this year. It was previewed on 24th April, 2026, and a lot has happened since then: GA releases, a new V4.1 Flash model, and a pricing table that looks different from what most older articles will provide you. This article on DeepSeek V4 consists of all the latest information about DeepSeek.
Key Importance:
DeepSeek V4 is available in two variations: V4-Pro and V4-Flash. V4-Flash is no longer available, but it is now being replaced with V4.1 Flash.
Both models are able to use 1M-token contexts and up to 384K-token outputs.
The weights are licensed under MIT, and the API is in the same format as the one set up by OpenAI.
API prices are a small fraction of what GPT-5.5 or Claude Sonnet charges, but they now differ between peak and off-peak hours.
Huawei says its Ascend chips fully support the V4 series.
What Is DeepSeek V4?
DeepSeek is a Chinese AI lab that became widely known in early 2025 with its R1 reasoning model. V4 is its next generation, built as a mixture-of-experts model family. That means only a fraction of the parameters are active for each token, which is how DeepSeek keeps running at low cost.
There are two main models. V4-Pro is the large one, with 1.6 trillion total parameters and 49 billion active. V4-Flash is the smaller, cheaper option with 284 billion total and 13 billion active parameters. DeepSeek says V4-Pro trails only the latest Gemini model on a world-knowledge benchmark, which is a strong claim, though it comes from the company itself.
A realistic caveat: DeepSeek's own technical report reportedly accepts that V4 sits a few months behind the very top closed models on some tasks. For most business work, that gap matters less than the price difference.
V4-Pro, V4-Flash and V4.1 Flash: What Changed Since Launch
The release history is easy to lose track of, so here it is in order:
April 24, 2026: V4-Pro and V4-Flash launch in preview.
July 31, 2026: V4-Flash-0731 becomes the official release, and the open weights go up on Hugging Face.
August 2026: V4-Pro-0813 follows as the official Pro release.
On 10 September 2026, V4.1 Flash is released, and the API's name is changed back to V4.1 Flash.
A planned September 14 change that would have routed V4-Pro requests to V4.1 Flash was later reversed, and V4-Pro stays on the API. Before hard-coding anything, have a look at the official model list; see how model names have changed so much this year.
One more thing worth clearing up: you may see articles about a "DeepSeek V4 Lite". DeepSeek's current lineup is V4-Pro and Flash, so treat Lite references as rumours from before launch.
Is DeepSeek Available in English?
Yes. It is the most often searched question about the model, which is why it has its own answer. The DeepSeek chatbot and the API are both in English, and V4 can process the enquiry and provide responses in English. It's easy to use even if you don't know Chinese. Communicate with the official DeepSeek website or via the API in English.
DeepSeek V4 API Pricing in 2026
The cheap $0.14 and $0.28 figures that circulated earlier belonged to the first Flash release. Today DeepSeek bills at two rates, and the off-peak rate is half the peak rate. These numbers are per million tokens, from the official pricing page as reported in September and October 2026:
Model | Peak input / output | Off-peak input / output |
V4.1 Flash | $0.30 / $1.20 | $0.15 / $0.60 |
V4-Pro | $1.32 / $3.96 | $0.66 / $1.98 |
Cached input is much cheaper still, around $0.003 per million tokens on V4.1 Flash off-peak. Prices and the off-peak window can change, so confirm both on the official DeepSeek API docs before you budget.
For scale, a price tracker puts GPT-5.5 at $5 input and $30 output, and Claude Sonnet at $3 and $15. On the price list, V4.1 Flash off-peak comes out roughly 95 percent cheaper or more. That is why teams with heavy, repetitive workloads keep testing DeepSeek.
How to Use the DeepSeek V4 API?
The API follows the OpenAI format, so most existing code only needs a new base URL and key. The base URL is https://api.deepseek.com, and the current model names are deepseek-flash for V4.1 Flash and deepseek-v4-pro for V4-Pro. The older deepseek-chat and deepseek-reasoner aliases are being phased out, so new projects should avoid them.
The API also offers an Anthropic-style endpoint and the Responses format, with tool calling. Both models support a thinking mode and a non-thinking mode. If you are choosing between them, start with Flash. It is far cheaper, and DeepSeek says its own tests put V4.1 Flash ahead of V4-Pro on agent benchmarks.
Open Weights and Self-Hosting
Both V4 models are released under the MIT license, with weights available on Hugging Face. In practice, that means you can run them on your own infrastructure. A 1 million token context and models this size do need serious hardware, so self-hosting is realistic for companies with GPU capacity, not for a laptop. If you are weighing local options at a smaller scale, our overview of open models can run better than V4 itself.
DeepSeek V4 and Huawei Ascend Chips
This is the part that made V4 a geopolitical story. Huawei said its entire Ascend supernode product line supports the V4 series and that its chips were used in part of the training process. DeepSeek's technical paper says it validated its approach on both Nvidia GPUs and Ascend chips, but DeepSeek has not said which hardware did the bulk of the training.
V4 runs on Huawei hardware, and that matters, but it is not proof that Nvidia is out of the picture. The Register's coverage is a detailed article on DeepSeek’s newest releases. DeepSeek also said Pro capacity is limited by compute constraints and that Pro pricing should fall once. Ascend 950 systems are deployed at scale. For the hardware side of the story, read our comparison of Huawei CANN vs CUDA.
Risks Businesses Should Weigh
Cheap tokens are only half the decision. Companies in the EU need to look at data protection and security before sending anything sensitive to a hosted API. German authorities have already pushed to pull the DeepSeek app from app stores, which we covered in our report on Germany's move to ban DeepSeek. Security researchers have also shown that DeepSeek models can be talked into unsafe behavior, which we explain in our piece on the DeepSeek jailbreak.
If you want the benefits without sending data to DeepSeek's servers, the open weights give you a second route: host the model yourself, or use a provider located in a region you trust. If you are comparing it with the earlier open model, our write-up on DeepSeek V3.1 shows how far the line has moved.
Who Should Use DeepSeek V4?
It fits teams with high-volume work where cost per token decides the budget: document processing, coding assistants, internal agents, and batch jobs that can run off-peak. It fits less well where you need strict data residency, the very strongest closed model on a hard task, or vendor guarantees you can put in a contract. Many teams end up using both, with DeepSeek for volume and a closed model for the hard cases.
Conclusion
DeepSeek V4 is worth a serious test if token cost is your main constraint. It is cheap, open, and easy to plug into existing code. It is not a free pass on privacy or security, and its prices and model names have changed several times already this year, so verify everything against the official docs before committing.
Need help choosing between DeepSeek, Claude, OpenAI, and open models for your business? Talk to TechNow, and we will help you pick, test, and deploy the right AI setup.
Frequently Asked Questions
What is DeepSeek V4?
It is DeepSeek's 2026 model family, built as a mixture-of-experts system with a 1 million token context window. It comes as V4-Pro and V4-Flash, and Flash has since been updated to V4.1 Flash.
How much does the DeepSeek V4 API cost?
V4.1 Flash costs $0.30 input and $1.20 output per million tokens at peak, and half that off-peak. V4-Pro costs $1.32 and $3.96 at peak, and half that off-peak.
Is DeepSeek V4 open source?
The weights are open under the MIT license and available on Hugging Face. The DeepSeek web and mobile apps are not open.
Is DeepSeek available in English?
Yes. The chatbot and API both work in English.
Does DeepSeek V4 run on Huawei chips?
Huawei says its Ascend supernode line supports the V4 series. The models were also validated on Nvidia GPUs.
Is the DeepSeek API compatible with the OpenAI SDK?
Yes. Point the SDK at https://api.deepseek.com, use your DeepSeek key, and choose a current model name.
How long is the DeepSeek V4 context window?
One million tokens, with up to 384K tokens of output.