DeepSeek R1 670B
#8 en Modèles de reasoningDeepSeek · 670b · 4× · vu le 27 août 2026
48
Momentum
DeepSeek R1 (670B/671B) is an Open Source Reasoning language model from DeepSeek released in January 2025. It is based on a Mixture-of-Experts architecture with 671 billion total parameters, of which approximately 37 billion are active per Token, and was trained primarily using Reinforcement Learning to develop Reasoning capabilities. The model achieves performance comparable to OpenAI o1 on mathematics and coding Benchmarks and is provided under the MIT license with open weights. Access is available through the DeepSeek API (model name "deepseek-reasoner"), chat interface, and Hugging Face.
Historique du momentum
01.06.30.08.
Fonctionnalités
| Key Benchmark (%) | AIME 2024: 79.8% Pass@1; MATH-500: 97.3% Pass@1 |
| Context Window (Tokens) | 64,000 tokens |
| License | MIT License (open weights, free commercial use and distillation) |
| Multimodality | No – text-only model, no image input/vision support |
| Platform | DeepSeek API (deepseek-reasoner), web chat (chat.deepseek.com), Hugging Face, iOS/Android app |
| Price per 1M Tokens | $0.55 input / $2.19 output (official DeepSeek API, as of 2026) |
| Release Date | January 20, 2025 |