DeepSeek Releases MIT-Licensed 685B DeepSeek-R1 Model Rivaling OpenAI's o1 at 30x Lower Cost

Jan 20, 2025, 12:34 PM

DeepSeek has officially released its new open-source reasoning models, DeepSeek-R1 and DeepSeek-R1-Zero, licensed under the MIT License. The models, boasting 685 billion parameters, perform on par with OpenAI's o1 model across math, code, and reasoning tasks. DeepSeek-R1 achieved 71.0% pass@1 on AIME 2024, comparable to OpenAI's o1, and reached 86.7% with majority voting, surpassing OpenAI's o1. The release includes a technical report detailing a novel training pipeline that uses reinforcement learning without supervised fine-tuning, incorporating a 'language consistency reward' to improve reasoning outputs. The models are available on Hugging Face, the DeepSeek website, and API, as well as in their chat and Android/iOS apps. DeepSeek also released distilled smaller models, including one based on Qwen-1.5B that outperforms GPT-4o and Claude-3.5-Sonnet on math benchmarks with 28.9% on AIME and 83.9% on MATH. The models offer significant cost savings, with DeepSeek-R1 being up to 30 times cheaper than OpenAI's o1.