DeepSeek has released an open-weights preview of its fourth-generation large language model, V4, claiming it slashes inference costs compared to its predecessor R1 while rivaling top proprietary American LLMs in performance. The model, now available in preview, notably runs on Huawei Ascend NPUs, expanding access for developers in regions affected by US export restrictions.
Catch: Inference cost reduction claims are based on DeepSeek's own internal benchmarks and have not yet been independently replicated by third-party evaluators.
Support for non-Nvidia hardware marks a significant shift, democratizing advanced AI capability and reducing dependence on a single chip supplier. The move could reshape the competitive landscape for open-source AI models.