De
DeepSeek
DeepSeek
Language modelSince 2024
A Chinese lab whose open, efficient reasoning models reset price/performance expectations.
Version timeline as of Jun 25, 2026.
Version timeline
7 versions- DeepSeek V4 FlashActivedeepseek-v4-flash· 1M ctx · cutoff Jan 1, 2026 · Fast
Budget/high-volume sibling of V4; 284B total / 13B active MoE; 1M context; MIT license; $0.14/$0.28 per 1M.
- deepseek-v4-pro· 1M ctx · cutoff Jan 1, 2026 · Flagship
Current flagship; 1.6T total / 49B active MoE, 384K max output, optional thinking mode that subsumes R1; 80.6% SWE-bench Verified; MIT license; $0.435/$0.87 per 1M.
Releases are dated and source-credited where available. Catch Me Up — everything that changed across AI & smart home ↗