DeepSeek V4
used in
0
stacks
DeepSeek's V4 generation of open-weight MoE models with 1M-token context and agent-focused reasoning modes.
DeepSeek V4 ships as two open-weight mixture-of-experts variants under the MIT license, with weights on Hugging Face. Per the model card, V4-Pro has 1.6T total parameters (49B activated) and V4-Flash has 284B (13B activated), both with a context length of one million tokens and three selectable reasoning effort modes. DeepSeek positions the release around stronger agent capabilities, and it is available on web, app, and the first-party API as well as via vLLM and SGLang.
Strengths
- MIT-licensed open weights for both Pro and Flash variants
- 1M-token context window
- Selectable reasoning effort modes (non-think plus two thinking modes)
- Sparse MoE design keeps activated parameters low relative to total size
Tradeoffs
- Branded as a preview release on deepseek.com as of June 2026
- V4-Pro's scale makes self-hosting impractical without large GPU clusters