DeepSeek rolls out V4 models

DeepSeek has released preview versions of DeepSeek-V4-Pro and DeepSeek-V4-Flash, its latest open-source language models, across its web app, mobile app and API. (deepseek.com) (cnbc.com) A language model is software that predicts the next chunk of text, then keeps going; newer versions are judged on how well they answer questions, write code and use tools. DeepSeek says V4 improves knowledge, reasoning and “agent” behavior, meaning the model can carry out multi-step tasks with external tools. (cnbc.com) (techxplore.com) The two V4 variants are aimed at different trade-offs. DeepSeek’s API docs list both models with a 1 million-token context window, support for tool calls and maximum output of 384,000 tokens. (api-docs.deepseek.com) DeepSeek’s pricing page lists V4-Flash at $0.14 per 1 million input tokens on a cache miss and $0.28 per 1 million output tokens. V4-Pro is listed at $1.74 per 1 million input tokens and $3.48 per 1 million output tokens, with a temporary 75% discount through May 5, 2026. (api-docs.deepseek.com) For developers, the launch also changes the product map. DeepSeek said on April 24 that API users should switch model names to `deepseek-v4-pro` or `deepseek-v4-flash`, while the older `deepseek-chat` and `deepseek-reasoner` aliases are set to stop working on July 24, 2026. (api-docs.deepseek.com) The company is also pitching V4 as easier to plug into existing software stacks. DeepSeek’s update says the new models work through OpenAI-format Chat Completions and an Anthropic-format endpoint without changing the base URL. (api-docs.deepseek.com) This release lands more than a year after DeepSeek’s R1 reasoning model rattled markets with claims of strong performance built on cheaper hardware. CNBC reported that analysts now see V4 less as a market shock than as another sign that Chinese AI models are getting cheaper and more capable. (cnbc.com) Reuters, via U.S. News, reported that DeepSeek has also adapted V4 to run on Huawei chips, reducing reliance on Nvidia and fitting Beijing’s push for a more self-sufficient artificial intelligence stack. (money.usnews.com) (techxplore.com) The near-term test is whether V4’s lower prices and long context window win developers before the July 24 API cutoff for older model names. DeepSeek has moved the launch from teaser to product, and now has to keep the service stable enough for people to build on it. (api-docs.deepseek.com) (status.deepseek.com)