DeepSeek V4 to Launch on Huawei Chips With One Trillion Parameters

DeepSeek is preparing to release V4, its most ambitious model yet, in the final weeks of April 2026. The one-trillion-parameter mixture-of-experts model will run on Huawei's domestically produced Ascend 950PR chips — a significant milestone in China's push for AI self-sufficiency.

Architecture and Performance

While DeepSeek V4 has one trillion total parameters, the MoE architecture means only approximately 37 billion activate per response. This design allows the model to run with the speed and efficiency of a 37B model while having access to the knowledge encoded in a trillion parameters.

Key specifications include a one million token context window — putting it in the same class as Google's Gemini models — and an 81 percent score on SWE-bench, suggesting strong performance on software engineering tasks. Pricing is projected at $0.30 per million tokens. The model will also be multimodal, supporting picture, video, and text generation.

The Huawei Connection

Reuters confirmed on April 4 that DeepSeek V4 will run on Huawei's Ascend 950PR chips, marking a departure from the NVIDIA hardware that most frontier models rely on. DeepSeek has been working with both Huawei and Cambricon to optimize the model for Chinese-made AI chips, a strategic imperative given US export restrictions on advanced semiconductors.

The decision to run on domestic hardware represents both a technical achievement and a political statement. If DeepSeek V4 delivers competitive performance on Chinese chips, it would undermine the premise that US export controls can bottleneck China's AI development.

Open Source Under MIT License

True to form, DeepSeek will release V4 under the MIT license, making it freely available for anyone to download, modify, and deploy. The training cost is estimated at just $5.2 million — a fraction of what US frontier labs spend — though this figure has not been independently verified.

Caveats

It is worth noting that the benchmark numbers cited come from internal testing and have not been independently verified as of publication. The AI research community will be watching closely for third-party evaluations once the model is publicly available.