DeepSeek R1
#2 v Reasoning modelyDeepSeek · od 20. Januar 2025 (DeepSeek-R1); Update DeepSeek-R1-0528 am 28. Mai 2025 · 106× · naposledy 17. 8. 2026
DeepSeek R1 is an open-weight reasoning model released by DeepSeek on January 20, 2025, using a Mixture-of-Experts architecture (671B total parameters, ~37B activated per token). It relies on large-scale reinforcement learning and explicit chain-of-thought output to achieve performance comparable to OpenAI-o1 on math, coding, and logic tasks. The model is released under an MIT license and is accessible via the DeepSeek API (model name deepseek-reasoner), the DeepSeek chat app (DeepThink mode), and third-party providers such as AWS Bedrock, Azure, Groq, and Hugging Face. A refreshed checkpoint (DeepSeek-R1-0528) improved accuracy and reduced hallucinations in May 2025.
Vlastnosti
| Key-Benchmark (%) | AIME 2024: 79,8%; MATH-500: 97,3% (Pass@1); AIME 2025 nach 0528-Update: 87,5% |
| Kontextfenster (Token) | 128.000 Token (max. Generierungslänge 32.768–64.000 Token) |
| Lizenz | MIT-Lizenz (Modellgewichte und Code, kommerzielle Nutzung erlaubt) |
| Multimodalität | Nein – reines Text-Modell, primär für Sprache, Mathematik und Code |
| Plattform | DeepSeek API (deepseek-reasoner), DeepSeek-Chat (DeepThink), Hugging Face, GitHub, AWS Bedrock, Azure, Groq, NVIDIA NIM |
| Preis pro 1M Token | Input: $0,14 (Cache-Hit) / $0,55 (Cache-Miss); Output: $2,19 (offizielle DeepSeek-API) |
| Release-Datum | 20. Januar 2025 (Original-Release); Update R1-0528 am 28. Mai 2025 |