DeepSeek V4’s official version may be launched by the end of this month, completing CUDA-free adaptation

PANews July 21 news, according to PConline Pacific Computer Technology, the DeepSeek-V4 official release has started a small-scale phased rollout test, including two versions: DeepSeek-V4-Pro and DeepSeek-V4-Flash. An insider said the official version is expected to be released by the end of this month, delayed compared with the previously planned mid-July. The new version will focus on optimizing reasoning efficiency, response speed, and agent capabilities. Developers participating in the test said that V4’s overall capabilities are benchmarked against Anthropic Opus 4.8; its code generation is close to OpenAI GPT-5.6 Sol, and its performance has improved in scenarios such as 3D modeling, SVG image generation, and game development. DeepSeek-V4 has completed the de-CUDA adaptation; on the first day after launch, it can already be compatible with multiple domestic AI chips, and it will use peak-and-valley time-of-use API billing, with peak-hour pricing at twice that of off-peak hours.