DeepSeek V4 is poised for an official release by the end of July, following a delay from its initial mid-July target. The new version, currently undergoing small-scale testing, includes DeepSeek-V4-Pro and DeepSeek-V4-Flash. It promises significant improvements in inference efficiency, response speed, and agent capabilities. The updated DeepSeek-V4 benchmarks against Anthropic Opus 4.8 and rivals OpenAI GPT-5.6 Sol in code generation. It also enhances performance in 3D modeling, SVG image generation, and mini-game development. Notably, DeepSeek-V4 has completed its CUDA-free adaptation, ensuring compatibility with various domestic AI chips. The platform will implement a peak/off-peak API billing system, with peak-hour rates set at double the off-peak pricing.