DeepSeek launches Huawei‑powered V4

A large language model is software trained on huge text datasets to predict the next word, which lets it answer questions, write code, and summarize documents. DeepSeek released preview versions of its new V4 models on April 24, with Huawei chip support built into the rollout. (techxplore.com) (api-docs.deepseek.com) DeepSeek named the two models DeepSeek-V4-Pro and DeepSeek-V4-Flash. Its API changelog says both became available on April 24, and the older `deepseek-chat` and `deepseek-reasoner` names will be discontinued on July 24, 2026. (api-docs.deepseek.com 1) (api-docs.deepseek.com 2) The company says V4-Pro is built for harder reasoning and coding tasks, while V4-Flash is the lower-cost, faster variant. DeepSeek’s documentation and outside coverage say the models support a 1 million-token context window, which is the amount of text they can keep in view in one session. (api-docs.deepseek.com) (techxplore.com) Huawei’s role is the part that shifts this from a normal model launch into an infrastructure story. Reuters reporting, carried by TechXplore and others, said Huawei chips were used in part of V4-Flash training, and Huawei said its Ascend systems fully support the V4 models. (techxplore.com) (thestar.com.my) That matters because Nvidia’s chips and its CUDA software stack still dominate advanced artificial intelligence training. DeepSeek’s earlier V3 and R1 models were trained on Nvidia hardware, while Business Today says V4 is the first flagship release the company optimized for Huawei’s domestic chip stack. (businesstoday.in) (techxplore.com) DeepSeek paired the launch with aggressive pricing. Its pricing page says V4-Pro is discounted 75% until May 5, 2026, and Reuters reported the company also cut input cache-hit prices across its lineup to one-tenth of the original level. (api-docs.deepseek.com) (money.usnews.com) Business Today reported V4-Pro at $1.74 per million input tokens and $3.48 per million output tokens, with V4-Flash positioned even lower. That price pressure is part of DeepSeek’s pitch against OpenAI, Anthropic, and Google as much as the benchmark claims are. (businesstoday.in) (api-docs.deepseek.com) DeepSeek says V4 improves knowledge, reasoning, and “agentic” work, meaning multi-step tasks the model can carry out with tools. TechXplore, citing DeepSeek’s own evaluations, reported the company claims V4-Pro trails OpenAI’s GPT-5.4 and Google’s Gemini 3.1-Pro only marginally on some reasoning benchmarks. (techxplore.com) Independent analysts are more careful than the launch materials. Omdia analyst Lian Jye Su told TechXplore the model appears competitive, while engineer Daniel Dewhurst told Reuters, via TechWire Asia, that wider testing is still needed before firm conclusions about performance. (techxplore.com) (techwireasia.com) The bigger point is that DeepSeek is no longer selling only a model. It is testing whether a Chinese software company can ship frontier-style artificial intelligence on Chinese chips, at Chinese prices, without depending as heavily on Nvidia’s hardware path. (techxplore.com) (businesstoday.in)