DeepSeek released DeepSeek-R1, an open-weight reasoning model that topped OpenAI's o1 on AIME 2024, MATH-500, and SWE-Bench Verified benchmarks1. The model's code and weights are licensed freely for commercial and personal use, including training new models on DeepSeek-R1 outputs.
DeepSeek-R1 is a mixture-of-experts transformer built on DeepSeek-V3-Base, carrying 671 billion total parameters with 37 billion active at any given time, and processing up to 128,000 tokens of input context. The model was fine-tuned over four stages to enhance its chain-of-thought reasoning, executing long lines of reasoning before producing output.
API access to DeepSeek-R1 costs $0.55 per million input tokens, $0.14 for cached inputs, and $2.19 per million output tokens. ANALYSIS That pricing positions DeepSeek-R1 well below typical frontier-model API rates for a model claiming benchmark parity with o1.
Beyond the o1 comparisons, DeepSeek-R1 outperformed Anthropic's Claude 3.5 Sonnet on 19 of 21 benchmarks and OpenAI's GPT-4o on 20 of 21 benchmarks. On LiveCodeBench, DeepSeek-R1 solved 65.9 percent of problems correctly.
DeepSeek also released seven related models alongside DeepSeek-R1. DeepSeek-R1-Zero follows a similar architecture but was fine-tuned entirely using reinforcement learning. Six additional dense models span parameter counts of 1.5 billion, 7 billion, 8 billion, 14 billion, 32 billion, and 70 billion.
The distilled variants posted strong results of their own. DeepSeek-R1-Distill-Qwen-32B outperformed OpenAI o1-mini across all benchmarks tested. DeepSeek-R1-Distill-Llama-70B beat OpenAI o1-mini on all benchmarks tested except Codeforces.
The mixture-of-experts architecture underpinning DeepSeek-R1 routes different inputs to different subsets of parameters via a gating module, allowing different expert networks to specialize in different types of examples. Because not all parameters are used for any given output, the model uses less energy and runs faster than dense models of comparable total size.
DeepSeek-R1-lite-preview, a smaller predecessor with fewer parameters and a different base model, was released in November.
DeepSeek has continued active model development since DeepSeek-R1's release. The company launched DeepSeek V4-Flash-0731 into public beta on July 31 with 304 billion total parameters and enhanced agentic capabilities ctx.
ANALYSIS The open licensing of DeepSeek-R1's weights — including permission to distill from its outputs — removes a barrier that has kept most frontier reasoning models proprietary. The distilled variants' performance against OpenAI o1-mini demonstrates that the reasoning capabilities transfer to substantially smaller architectures.