DeepSeek R1

2025-01-20 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases R1, a 671-billion-parameter reasoning model (37 billion active) under the MIT license, whose announced scores match those of OpenAI o1, along with six distilled models.

### Context At the end of 2024, long chain-of-thought reasoning models were the preserve of closed offerings, with OpenAI o1 as the reference. DeepSeek had just released V3 (December 2024), an open general-purpose model, and R1 is its reasoning-specialized variant. The stakes of this release are twofold: bring an open model closer to o1's performance, and let everyone reuse the model and its outputs, including to distill smaller models. ### What the model brings - Performance announced by the publisher as "on par with OpenAI-o1" in mathematics, code and reasoning, obtained through large-scale reinforcement learning in post-training. - Weights and a technical report fully published under the MIT license: commercial use, modification and distillation allowed. - An explicit license change: the model's outputs can be used to train other models. - Six distilled models (Qwen 1.5B, 7B, 14B, 32B and Llama 8B, 70B), of which the 32B and 70B variants are announced at the level of OpenAI o1-mini. - Immediate access on the website (DeepThink button) and through the API under the identifier `deepseek-reasoner`. ### Architecture and specifications - Type: Mixture-of-Experts, 671 billion parameters in total, 37 billion activated per token. - Context: 128K tokens (model card on Hugging Face). - Modalities: text. Number of experts, training data and cutoff date: not disclosed in the sources consulted. - License: MIT for the code and the weights, commercial use allowed. The distilled models inherit the licenses of their bases (Apache 2.0 for the Qwen2.5 bases, Llama licenses for the Llama bases). - Settings recommended by the publisher: temperature between 0.5 and 0.7 (0.6 advised), no system prompt, instructions placed in the user message. - Weights: huggingface.co/deepseek-ai/DeepSeek-R1. ### Cost and access | `deepseek-reasoner` API (USD per million tokens) | Price | |---|---| | Input, cache hit | 0.14 | | Input, cache miss | 0.55 | | Output | 2.19 | Hardware required for self-hosting: not disclosed in the sources consulted. ### Benchmarks | Benchmark | Claude-3.5-Sonnet | GPT-4o 0513 | OpenAI o1-mini | OpenAI o1-1217 | DeepSeek-R1 | |---|---|---|---|---|---| | MMLU (Pass@1) | 88.3 | 87.2 | 85.2 | 91.8 | 90.8 | | MMLU-Pro (EM) | 78.0 | 72.6 | 80.3 | not published | 84.0 | | GPQA Diamond (Pass@1) | 65.0 | 49.9 | 60.0 | 75.7 | 71.5 | | AIME 2024 (Pass@1) | 16.0 | 9.3 | 63.6 | 79.2 | 79.8 | | MATH-500 (Pass@1) | 78.3 | 74.6 | 90.0 | 96.4 | 97.3 | | Codeforces (rating) | 717 | 759 | 1820 | 2061 | 2029 | | LiveCodeBench (Pass@1) | 33.8 | 34.2 | 53.8 | 63.4 | 65.9 | Source: DeepSeek-R1 model card on Hugging Face, scores measured and published by the publisher. ### Versus closed models At the time of its release, the closed reference model compared by the publisher was OpenAI o1-1217. In the table, R1 is ahead of o1-1217 on AIME 2024 (79.8 versus 79.2), MATH-500 (97.3 versus 96.4) and LiveCodeBench (65.9 versus 63.4). It remains behind on MMLU (90.8 versus 91.8), GPQA Diamond (71.5 versus 75.7) and Codeforces (2029 versus 2061). It clearly outperforms GPT-4o and Claude 3.5 Sonnet on mathematical reasoning tasks. On price, the publisher lists 2.19 dollars per million output tokens, an open API rate that the post sets against OpenAI's offering in a visual not reproduced here. The open choice therefore becomes credible for reasoning: downloadable weights, MIT license and the possibility of distillation. ### Key takeaways R1 is the first open-weights reasoning model presented at o1's level, with a license free of commercial restrictions. It comes with six distilled models that can be used on much more modest hardware than the full model.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Raisonnement, Distillation