Woofun AI reports that DeepSeek-V4 has initiated small-scale gray testing with Pro and Flash variants, targeting an official launch by month-end. The update prioritizes inference efficiency, response speed, and agent capabilities, while completing adaptation to remove CUDA dependency for compatibility with multiple domestic AI chips.

Testing participants indicate that V4's comprehensive performance matches Anthropic Opus 4.8, with code generation nearing OpenAI GPT-5.6 Sol levels. Improvements are noted in 3D modeling, SVG image generation, and mini-game development. The release will implement a peak and off-peak API billing structure, where peak rates are double the off-peak cost.