DeepSeek-V4 Model Released Open-Source, Huawei Cloud First to Adapt, Enabling 1M Token Long Context Accessibility

The DeepSeek-V4 model has been officially released and open-sourced, with Huawei Cloud being the first to adapt it. The model features a 1M-token ultra-long context and leads in Agent capabilities, world knowledge, and reasoning performance in domestic and open-source domains.

The DeepSeek-V4-Flash variant has reduced parameters to 284B, further lowering inference costs. It offers faster and more cost-effective API services, making 1M-token context accessible to more users. Huawei Cloud's MaaS platform provides developers with deployment-free, one-click access to the model's API.

Huawei Cloud collaborated across system, operator, and cluster layers, using layered attention compression, Ascend high-performance fusion operators, and framework optimizations to ensure rapid adaptation and high-performance inference for native 1M-token long context.