DeepSeek launches V4 model
- DeepSeek released a preview of its DeepSeek-V4 family on April 24, putting V4-Pro and V4-Flash on its website, app and API, and publishing model weights through Hugging Face. - DeepSeek says both models handle up to 1 million tokens; V4-Pro is priced at $1.74 per 1 million input tokens before a temporary 75% discount, while V4-Flash starts at $0.14. - The launch ties DeepSeek’s next model cycle to Huawei chips and open distribution, extending China’s push to build advanced AI systems with less reliance on Nvidia. (reuters.com)
DeepSeek has released a preview of its DeepSeek-V4 family, adding V4-Pro and V4-Flash across its website, mobile app and developer API. (deepseek.com) (api-docs.deepseek.com) Large language models predict the next word in a sequence, and the cost that matters to customers is often inference, the price of generating answers after training is done. DeepSeek’s pricing page lists V4-Flash at $0.14 per 1 million input tokens and V4-Pro at $1.74 before a temporary 75% discount. (api-docs.deepseek.com) DeepSeek says both V4 models support a 1 million-token context window, which is the amount of text a model can keep in view at once. Its API documentation also says the older `deepseek-chat` and `deepseek-reasoner` names will be deprecated on July 24, 2026, in favor of the new V4 line. (api-docs.deepseek.com 1) (api-docs.deepseek.com 2) The company is distributing the models beyond its own platform. DeepSeek’s Hugging Face collection shows DeepSeek-V4-Pro, DeepSeek-V4-Flash and corresponding base models, with V4-Pro-Base listed at 1.6 trillion parameters and V4-Flash-Base at 292 billion. (huggingface.co 1) (huggingface.co 2) Reuters reported on April 24 that DeepSeek adapted V4 for Huawei chip technology, a shift from its earlier dependence on Nvidia hardware. Huawei said its Ascend chips were used in part of V4’s training process. (reuters.com) That hardware detail lands in the middle of a wider fight over export controls and China’s push for domestic substitutes. Reuters quoted Omdia analyst He Hui saying support for DeepSeek V4 shows top Chinese AI models can now run on Chinese hardware. (reuters.com) DeepSeek is also framing V4 as a stronger reasoning and agent model, not just a cheaper chatbot. Its homepage says the preview has “top-tier reasoning performance” and significantly improved agent capabilities. (deepseek.com) The immediate test is whether developers move from trying V4 to building on it. DeepSeek has already put the models where that decision gets made: the API, the app, the web client and Hugging Face. (deepseek.com) (api-docs.deepseek.com) (huggingface.co)
