DeepSeek adapts V4 for Huawei chips
- Chinese startup DeepSeek released a preview of DeepSeek-V4 on Friday, saying the new open-source model is adapted for Huawei Ascend chips and offered in Pro and Flash versions. - Reuters reported Huawei chips were used in part of V4’s training, while DeepSeek said V4-Pro beats other open-source models on world-knowledge benchmarks and trails only Google’s Gemini-Pro-3.1. - The launch shifts DeepSeek further from Nvidia and adds to China’s push to pair domestic models with domestic chips. (reuters.com)
DeepSeek released a preview of its new V4 model on Friday and said it is adapted to run on Huawei’s Ascend chips. (reuters.com) The Hangzhou-based startup offered V4 in two versions, Pro and Flash, and open-sourced the release so developers can download and modify it. (cnbc.com) (reuters.com) Reuters reported Huawei said its chips were used in some of V4’s training process, a change from DeepSeek’s earlier reliance on Nvidia hardware. (reuters.com) A large language model is software that predicts the next word over and over; the chip underneath it is the engine that determines how fast and cheaply that work gets done. DeepSeek’s new release ties the model more directly to China’s best-known domestic AI chip line. (cnbc.com) (reuters.com) DeepSeek said V4-Pro outperforms other open-source models on world-knowledge benchmarks and trails only Google’s closed-source Gemini-Pro-3.1 on that measure. CNBC reported the company is also pitching V4 as strong on agent-based tasks, knowledge processing and inference. (reuters.com) (cnbc.com) Inference is the cost of running a trained model after it is built. Counterpoint Research’s Neil Shah told CNBC that V4 offers lower inference costs than earlier models, and Wei Sun said its benchmark profile suggests strong agent capability at lower cost. (cnbc.com) Huawei’s Ascend software stack is already publishing deployment instructions for DeepSeek-V4-Flash, including single-node setups on Atlas 800 A3 and A2 systems with eight accelerators per node. (docs.vllm.ai) Analysts framed the release as part of a broader hardware-and-model pairing strategy inside China. Omdia’s He Hui told Reuters that supporting DeepSeek V4 shows top Chinese AI models can now run on Chinese hardware. (reuters.com) The timing also lands in the middle of a sharper U.S.-China technology fight. Reuters said the launch came a day after the White House accused China of stealing United States AI labs’ intellectual property on an industrial scale. (reuters.com) DeepSeek’s last major shock came with R1 in January 2025, when its low reported training cost rattled assumptions about how much money frontier AI requires. V4 does not repeat that surprise, but it does show where DeepSeek wants its next stack to run. (cnbc.com) (reuters.com)
