DeepSeek Launches DeepSeek-V4 Preview
The preview version of DeepSeek's new series model, DeepSeek-V4, has officially launched and been open-sourced simultaneously. It includes two versions: deepseek-v4-pro and deepseek-v4-flash, both supporting 1M ultra-long context. V4-Pro matches top-tier closed-source models in Agent coding, world knowledge, and reasoning performance; V4-Flash provides comparable reasoning capabilities at a lower cost. The model adopts a new attention mechanism and DSA sparse attention, significantly reducing long context computation and memory overhead.
The preview version of DeepSeek's new series model, DeepSeek-V4, has officially launched and been open-sourced simultaneously. It includes two versions: deepseek-v4-pro and deepseek-v4-flash, both supporting 1M ultra-long context. V4-Pro matches top-tier closed-source models in Agent coding, world knowledge, and reasoning performance; V4-Flash provides comparable reasoning capabilities at a lower cost. The model adopts a new attention mechanism and DSA sparse attention, significantly reducing long context computation and memory overhead.
This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.