DeepSeek-V3
ConditionalDeepSeek · DeepSeek · Released 2024-12
671B total, 37B active (MoE)Text
A massive sparse MoE that matched or beat leading closed models on many benchmarks at a fraction of the reported training cost.
Strengths
- +Frontier-competitive reasoning, coding, and math
- +Only ~37B active parameters per token despite 671B total
- +Commercial use permitted under its model license
Limitations
- -671B total parameters require a multi-node GPU cluster to serve
- -Custom model license, not a standard OSI license like Apache/MIT
License
DeepSeek Model License (commercial use permitted)
free with acceptable-use or scale conditions - read it
Hardware
needs multi-GPU or datacenter hardware to run well
Links
Stats
deepseek-ai/DeepSeek-V3— downloads·— likes
via Hugging Face