DeepSeek R1: Open-Source Reasoning at Scale

DeepSeek has released R1, a 671-billion-parameter Mixture-of-Experts reasoning model with fully open weights. The model matches or exceeds the performance of OpenAI's o1 on several reasoning benchmarks while costing a fraction of the price to run.

R1 uses a distilled chain-of-thought approach, trained via reinforcement learning to produce transparent reasoning traces. On AIME 2024, R1 achieves a 79.8% solve rate. On GPQA Diamond, it scores 71.5%. The model also performs well on coding tasks, though it is primarily optimized for mathematical and logical reasoning.

What sets R1 apart is its openness. The full model weights are available under a permissive license, allowing researchers and companies to deploy it on their own hardware, fine-tune it for specific domains, and inspect its behavior at every level. This stands in contrast to OpenAI's o1, which remains closed-source and significantly more expensive.

Through API providers like OpenRouter, R1 is available at $0.55 per million input tokens, making it one of the most cost-effective reasoning models available. The open-source community has also produced smaller distilled variants that run on consumer hardware.

The release has accelerated the trend of open-source models competing at the frontier. Following R1's success, several other labs have released their own reasoning models with open weights, creating a competitive ecosystem that benefits developers and researchers.