DeepSeek model lab
MIT-licensed weights and codeDeepSeek
DeepSeek focuses on efficient mixture-of-experts systems, reasoning, coding and agentic tasks, with unusually permissive model releases.
Latest tracked open release
DeepSeek V4
DeepSeek V4 ships in Pro and Flash variants, extending the family to million-token context with stronger reasoning, coding and agentic performance.
- Released
- 24 April 2026
- Licence
- MIT
- Pro scale
- 1.6T total, 49B active parameters
- Flash scale
- 284B total, 13B active parameters
- Context
- 1M tokens
Why it matters
Where DeepSeek stands out
- Permissive MIT release
- Pro and efficient Flash variants
- Million-token context
Go to the source