DeepSeek-R1 open-sources RL recipe, distilled models

This is one of the rare drops that moves both research transparency and developer usability forward at the same time.

• –DeepSeek-R1-Zero shows pure RL can elicit advanced reasoning behaviors without an initial SFT stage, then DeepSeek-R1 adds cold-start and alignment stages to improve readability and stability.
• –The distilled Qwen/Llama variants turn frontier-style reasoning into deployable sizes, which matters more for real teams than a single flagship model.
• –DeepSeek reports parity or wins versus o1 on several math/coding benchmarks, and third-party Open-R1 reproductions broadly land in the same neighborhood with expected sampling variance.
• –Open licensing and released checkpoints lower the barrier for fine-tuning, self-hosting, and downstream experimentation across the open model ecosystem.