DeepSeek releases DeepSeek-V4 preview version

The preview version of DeepSeek's new model series, DeepSeek-V4, has been officially released and open-sourced. It includes two versions: deepseek-v4-pro and deepseek-v4-flash, both supporting ultra-long contexts of up to 1MB. V4-Pro rivals top-tier closed-source models in agent encoding, world knowledge, and inference performance; V4-Flash offers similar inference capabilities at a lower cost. The models employ a novel attention mechanism and DSA sparse attention, significantly reducing computational and memory overhead for long contexts.