DeepSeek
DeepSeek models you can run locally
DeepSeek builds open-weight reasoning and coding models, including efficient distilled variants and large mixture-of-experts systems.
Available models
Ordered from the lightest model to the largest.
DeepSeek R1 1.5B
1.5B · Dense
Tiny reasoning model distilled from R1
DeepSeek R1 Distill 7B
7B · Dense
R1 reasoning distilled into Qwen 7B
DeepSeek R1 Distill 14B
14B · Dense
R1 reasoning distilled into Qwen 14B
DeepSeek R1 Distill 32B
32B · Dense
R1 reasoning distilled into Qwen 32B — sweet spot
DeepSeek V4 Flash
284B · 13B active · MoE
Efficient long-context V4 — 13B active, 1M context
DeepSeek R1
671B · 37B active · MoE
Massive MoE reasoning model — 37B active
DeepSeek V3.2
685B · 37B active · MoE
State-of-the-art MoE — 37B active params
DeepSeek V4.1 Flash
763B · 16B active · MoE
Multimodal V4.1 MoE — 552B backbone plus 196B Engram memory, 16B active and a quarter of V4 Flash's KV cache
DeepSeek V4 Pro
1.6T · 49B active · MoE
Flagship V4 MoE, official 0813 release — 49B active, 1M context, DSpark speculative decoding