Landmark
DeepSeek-R1 Released
DeepSeek released R1, an open-source reasoning model that matched OpenAI o1 performance on many benchmarks at a fraction of the cost. The model was trained with only $6M in compute using reinforcement learning without supervised fine-tuning, challenging assumptions about AI development costs and strategies. DeepSeek-R1 achieved approximately 15.8% on the ARC-AGI benchmark, demonstrating strong abstract reasoning capabilities for an open-source model.
- Open-source reasoning model
- Matched OpenAI o1 performance
- Trained for ~$6M in compute
- Pure RL without SFT
- 15.8% on ARC-AGI benchmark
- Challenged AI cost assumptions