DeepSeek Releases R2 Reasoning Model

DeepSeek, the Chinese AI research lab funded by quantitative trading firm High-Flyer Capital, has released R2 — a reasoning-focused model that sets new records on mathematical and scientific benchmarks. R2 achieves 97.8% on the MATH benchmark and 95.3% on GPQA Diamond, surpassing all existing models including GPT-5.2 and Claude Opus 4.6 on these specific evaluations.

R2 builds on the techniques introduced in the original DeepSeek R1, which pioneered reinforcement learning approaches for developing reasoning capabilities in language models. The new model uses an expanded training process that includes multi-step mathematical proofs, scientific reasoning chains, and code execution verification, resulting in dramatically improved performance on complex analytical tasks.

The model is released in multiple sizes — 671B (full), 70B, and 7B parameters — with all versions available under an open-source license. The 7B version, in particular, has attracted significant attention for achieving reasoning performance comparable to GPT-4o while being small enough to run on consumer hardware.

DeepSeek's API pricing continues to undercut Western competitors significantly, with R2 available at $2 per million input tokens and $8 per million output tokens for the full 671B model. The pricing has prompted renewed discussion about the cost of AI services and the competitive dynamics between Chinese and American AI companies.

The release has significant geopolitical implications. Despite US export controls on advanced AI chips, DeepSeek has demonstrated that innovative training techniques can partially compensate for hardware limitations. R2 was reportedly trained using a combination of older NVIDIA A100 GPUs and domestically produced Huawei Ascend chips, challenging assumptions about the effectiveness of chip export restrictions.