DeepSeek adapts to Huawei

DeepSeek on April 24 released a preview of DeepSeek-V4, a new model line adapted to run on Huawei’s Ascend chips. (reuters.com) The launch included two versions, DeepSeek-V4-Pro and DeepSeek-V4-Flash, and DeepSeek said the models were open-sourced at release. Reuters reported the move marks a sharp contrast with DeepSeek’s earlier reliance on Nvidia hardware. (reuters.com) (techxplore.com) Huawei said its chips were used in some of V4’s training process. The company also said its Ascend supernode, built on Ascend 950 artificial intelligence chips, would fully support DeepSeek’s V4 versions. (reuters.com) (finance.yahoo.com) A large language model is software trained on huge amounts of text so it can predict the next word and generate answers, code, or summaries. The chips matter because training and running those systems requires dense, specialized computing hardware, and Nvidia has dominated that market. (cnbc.com) (reuters.com) DeepSeek said the Pro version outperformed other open-source models on world-knowledge benchmarks and trailed only Google’s closed-source Gemini-Pro-3.1 on that measure. It also said V4 was particularly suited to “agentic” work, meaning systems that can carry out multistep tasks with tools. (reuters.com) (techxplore.com) The timing reflects pressure created by U.S. export controls that have limited Chinese access to the most advanced American artificial intelligence chips. Analysts told Reuters the DeepSeek-Huawei tie-up shows those restrictions are also accelerating local substitutes in chips, systems software, and cloud infrastructure. (reuters.com) Counterpoint Research’s Neil Shah told CNBC the preview showed lower inference costs than prior models. Reuters separately reported that DeepSeek’s earlier low-cost releases had already unsettled global investors and pushed rivals to defend their pricing and performance. (cnbc.com) (reuters.com) South China Morning Post reported Huawei described the adaptation as “day zero,” meaning support was ready as V4 arrived rather than months later. That detail suggests the partnership was not just about compatibility after launch, but about joint preparation before release. (scmp.com) DeepSeek’s last big model, V3, was released in late 2024, and Reuters said V4 arrived about 15 months later. This time, the headline was not only the model itself, but the hardware underneath it. (techxplore.com) (reuters.com)