DeepSeek has released R1, an open-weights reasoning model that scores within a few points of the best proprietary systems on graduate-level math, competitive programming, and formal logic benchmarks. The release marks the clearest signal yet that the gap between open and closed AI systems is narrowing on tasks that require multi-step thinking.
Reasoning models differ from standard large language models in that they generate an internal chain of thought before producing a final answer. This approach, pioneered by OpenAI’s o1 series and refined by Google and Anthropic, dramatically improves performance on problems that require planning, backtracking, or verification. Until now, the most capable reasoning systems have been available only behind API paywalls.
Architecture and training
R1 uses a mixture-of-experts architecture with 671 billion total parameters, though only 37 billion are active for any given token. This design keeps inference costs manageable while preserving the model’s capacity for complex reasoning. DeepSeek trained R1 using large-scale reinforcement learning from verifiable rewards, a technique where the model is rewarded for producing correct final answers on math and code problems rather than for mimicking human-written explanations.
The team also released a series of distilled versions ranging from 1.5 billion to 70 billion parameters. These smaller models inherit much of R1’s reasoning ability and can run on consumer hardware, making them accessible to researchers and developers who lack access to large GPU clusters.
Implications for the industry
The release puts pressure on proprietary model providers to justify premium pricing for reasoning capabilities. If open alternatives can match performance at a fraction of the cost, enterprise buyers gain significant leverage in negotiations.
For the research community, R1’s open weights enable studies that are impossible with API-only models. Teams can now probe the internal representations that emerge during chain-of-thought reasoning, potentially unlocking insights into how and why these models succeed or fail on specific problem types.
DeepSeek has published a detailed technical report alongside the model weights, including ablation studies on the reinforcement learning process and analysis of failure modes. The model is available under a permissive license that allows commercial use.

