DeepSeek R1
#2 v Reasoning modelydeepseek · od 2025-01-20 · 81× · naposledy 03. 7. 2026
96
Momentum
DeepSeek R1 is an open-source reasoning model by DeepSeek, released on January 20, 2025. It is built on a Mixture-of-Experts (MoE) architecture with 671 billion total parameters, of which only approximately 37 billion are active per inference pass. The model uses reinforcement learning (RL) with chain-of-thought training and achieves benchmark performance comparable to OpenAI o1 in mathematics, coding, and reasoning tasks. It is released under the MIT license and is available both via the DeepSeek API and as open-weight model for self-hosting.
Vývoj momenta
04.04.03.07.
Vlastnosti
| Key Benchmark (%) | AIME 2024: 79.8% (Pass@1) · MATH-500: 97.3% (Pass@1) · MMLU: 90.8% (Pass@1) — from the official technical paper (arxiv.org/abs/2501.12948). Update R1-0528 (May 2025): AIME 2025: 87.5% |
| Context Window (Tokens) | 128,000 tokens (128K) – per technical report and model card; updated version R1-0528 approx. 130K tokens |
| License | MIT License (model weights & API outputs; commercial use, modifications, and distillation permitted) |
| Multimodality | No – DeepSeek R1 supports text input only (no image, audio, or video input). For multimodal tasks, DeepSeek offers a separate model line (Janus series). |
| Platform | DeepSeek API (api.deepseek.com, model=deepseek-reasoner) · DeepSeek Chat (chat.deepseek.com) · Open-source weights on Hugging Face & GitHub (self-hosting via vLLM, llama.cpp, Ollama, among others) |
| Price | API (DeepSeek official, at release): $0.14/1M input tokens (cache hit) · $0.55/1M input tokens (cache miss) · $2.19/1M output tokens |
| Price per 1M Tokens | Input (cache miss): $0.55 · Input (cache hit): $0.14 · Output: $2.19 (official DeepSeek API pricing at launch; deepseek-reasoner alias) |
| Release Date | January 20, 2025 |