DeepSeek has released R2, its latest reasoning-optimized large language model, which achieves top-tier performance on mathematical reasoning and competitive programming benchmarks while the lab claims it was trained using significantly less compute than comparable Western models. The release continues the pattern established by DeepSeek R1, which generated significant attention in early 2025 by demonstrating that strong reasoning capability did not require the scale of compute investment that US-based labs had applied.
The R2 model achieves competitive scores on MATH-500, AIME (American Invitational Mathematics Examination), and Codeforces competitive programming benchmarks, placing in the top tier of all published models on reasoning-intensive tasks. On some mathematics benchmarks, R2's published scores exceed those of Claude Sonnet 4 and GPT-5, though independent benchmark replication — which often reveals gaps between published and reproducible results — has not yet been completed at scale.
DeepSeek's training cost claims, if accurate, have significant implications for the semiconductor export control debate. US controls on exporting NVIDIA H100 and A100 chips to China were partly premised on the assumption that frontier AI capability required the largest NVIDIA clusters. DeepSeek's efficient training techniques — which include mixture-of-experts architectures, novel parallelism strategies, and aggressive data curation — challenge this assumption by demonstrating that competitive results can be achieved with less total compute and consequently on hardware not subject to the highest-tier export controls.
The R2 model is available for download as open weights for non-commercial use, and DeepSeek offers API access for commercial applications. The model runs on standard NVIDIA hardware available in regions without export controls, making it accessible to researchers and businesses in Europe and Asia without dependence on the US hardware supply chain.
Western AI labs have responded to DeepSeek's efficiency results by accelerating their own efficiency research, with several publishing papers on distillation techniques and training optimization methods that borrow from and extend the approaches DeepSeek has published. The competitive pressure from efficient training is changing the economics of frontier AI development in ways that benefit organizations outside the top tier of compute-rich US labs.