On April 24, the preview edition of DeepSeek-V4 was formally unveiled and made available as open-source software. This model has set a new benchmark in China and the wider open-source community, excelling in three key domains: agent capabilities, world knowledge, and reasoning prowess. Available in both Pro and Flash versions, the model supports ultra-long contexts, accommodating up to one million tokens. Notably, it substantially decreases computational and memory demands, with a remarkable 73% reduction in inference FLOPs and a significant 90% decrease in KV cache memory utilization.
