DeepSeek-V3
Frontier-class weights, with an asterisk worth reading
What it is
A 671-billion-parameter mixture-of-experts model (37B active per token) with multi-head latent attention and multi-token prediction — one of the strongest open-weight frontier-class models, distributed with paper and inference guidance.
Why it's interesting
A textbook case for reading licenses separately: the code is MIT, but these weights ship under the custom DeepSeek Model License — commercial use permitted, behavioral restrictions attached. Later DeepSeek releases moved to MIT weights; this checkpoint didn't.
Use cases
- Self-hosted frontier-class inference
- Research on MoE architectures
- Distillation into smaller models
Who it's for
Well-resourced infra teams and researchers
Setup
Advanced. Multi-GPU H100/H200-class hardware for full weights (hundreds of GB); most users consume via providers; SGLang, vLLM, or TensorRT-LLM to serve
Limitations & cautions
The weights license carries use restrictions that keep it outside strict OSI/OSAID definitions, and the hardware bill is enormous.
Editorial takeaway
Historically important and technically superb — and the exact reason this site prints weights licenses next to code licenses.