DeepSeek R1 671B
#24 in Reasoning-ModelleDeepSeek · 671b · 2× · tolest 23. Juli 2026
14
Momentum
DeepSeek R1 671B is an Open-Weight Reasoning model from Chinese AI lab DeepSeek based on a Mixture-of-Experts architecture with 671 billion total parameters, of which only around 37 billion are actively used per Inference step. The model was primarily trained using Reinforcement Learning and displays visible Chain-of-Thought Reasoning steps before the final answer. It is provided under the MIT license as an Open-Weight model and is additionally available through the official DeepSeek API as well as the chat service chat.deepseek.com. An update (DeepSeek-R1-0528) further improved Reasoning depth and Benchmark performance in May 2025.
Momentum-Verloop
19.05.17.08.
Features
| Key Benchmark (%) | AIME 2024: 79.8% Pass@1; MATH-500: 97.3% Pass@1 |
| Context Window (Tokens) | 163,840 tokens |
| License | MIT License, permits commercial use and modification |
| Multimodality | Not multimodal; text-only reasoning model (max generation length 32,768 tokens) |
| Platform | Open-weight download (Hugging Face/GitHub), chat at chat.deepseek.com, API via platform.deepseek.com |
| Price per 1M Tokens | $0.55 input / $2.19 output per 1M tokens (cache-hit input: $0.14) |
| Release Date | January 20, 2025 |