DeepSeek R1 matches the reasoning performance of OpenAI o1 at a fraction of the cost. This model uses a Mixture-of-Experts architecture that activates only 37 billion parameters out of 671 billion for each task.
Training and efficiency facts
DeepSeek R1 achieves its results through a multi-stage training pipeline. Developers used cold-start data consisting of thousands of long Chain-of-Thought examples before applying reinforcement learning. This process addresses readability issues seen in the precursor DeepSeek-R1-Zero. While R1-Zero relied on pure reinforcement learning to develop reasoning, it suffered from language mixing and poor readability. DeepSeek R1 fixes these problems by using a readable output format that includes a summary at the end of each response. The model also uses a language consistency reward to mitigate language mixing during training, though this reward causes a small performance degradation.
The model delivers high performance in math and coding tasks. DeepSeek R1 scored 79.8% on the AIME 2024 benchmark, while OpenAI o1 scored 78.2%. On the MATH-500 benchmark, DeepSeek R1 achieved a 97.3% pass rate compared to 96.4% for OpenAI o1. You should know that running the model locally via Ollama prevents your data from being used for training.
| Feature | DeepSeek R1 | OpenAI o1 |
|---|---|---|
| License | MIT | Proprietary |
| Math (AIME 2024) | 79.8% | 78.2% |
| Math (MATH-500) | 97.3% | 96.4% |
| Input Cost (per 1M tokens) | $1.37 | $75.00 |
Scaling and deployment realities
DeepSeek R1 is not a casual experiment from a small team. High-Flyer, a $7B hedge fund, backs the project. The development team includes members with Olympic-level medals in mathematics and physics. DeepSeek R1 operates with access to approximately 50,000 GPUs. Some claims suggest the training cost was only $5.5 million, but this figure excludes costs for ablations, experimental runs, and data curation.
Smaller distilled versions exist for local use. The R1-Distill-Qwen-32B model outperforms OpenAI o1-mini on several metrics. These distilled models use supervised fine-tuning on 800k samples from DeepSeek-R1 but do not include a reinforcement learning stage. However, medical analysis shows that longer reasoning responses in DeepSeek R1 correlate with a higher probability of error. In clinical cases, shorter responses under 5,000 characters associate more strongly with accuracy. Does the model’s tendency to overthink hinder its reliability in high-stakes environments?
