DeepSeek v4 Runs on Huawei Chips, No CUDA Required

DeepSeek v4 runs on Huawei's Ascend chips with zero CUDA dependency. That's a first for a competitive frontier model, and it validates an entire non-NVIDIA hardware stack, much like AMD's Lemonade.

The version 4 API landed today with two models: deepseek-v4-flash and deepseek-v4-pro. Both reportedly perform well on benchmarks like MMLU-Pro and SWE-bench Verified. The API is compatible with OpenAI and Anthropic formats, so developers can swap base URLs and keep using existing SDKs via platforms like OpenRouter.

The models are MIT-licensed and priced competitively. The older deepseek-chat and deepseek-reasoner names will be deprecated in July 2026, mapping to flash's non-thinking and thinking modes respectively.