DeepSeek R1 — Specifications
| Développeur | DeepSeek |
|---|
| Type | LLM (MoE, reasoning) |
|---|
| Modalité | Text → Text |
|---|
| Paramètres | 671B total / 37B active (MoE) |
|---|
| Fenêtre de contexte | 128K |
|---|
| Sortie maximale | — |
|---|
| Licence | MIT (open) |
|---|
| Poids ouverts | ✅ Yes |
|---|
| Date de sortie | 2025 |
|---|
| Input price | $0.50 /1M |
|---|
| Output price | $2.15 /1M |
|---|
| API providers | DeepSeek, DeepInfra, OpenRouter |
|---|
🖥️ Run it locally
| VRAM (4-bit) | ~400 GB |
|---|
| Minimum GPU | Multi-GPU server |
|---|
Official page →
DeepSeek’s landmark open reasoning model — a 671B mixture-of-experts (37B active), MIT-licensed, delivering frontier-class chain-of-thought reasoning at a fraction of proprietary cost. Server-class to self-host, but openly downloadable.
