DeepSeek releases the open-weight reasoning model R1
DeepSeek publishes R1, R1-Zero, six distilled models, and its reinforcement-learning recipe, pushing open reasoning models into the frontier conversation.
What happened
Event details
DeepSeek-R1 packaged a long-running research program into a broad release: the R1-Zero experiment, the production R1 model, six smaller distilled checkpoints, a technical paper, web access, and an API. The main release used reinforcement learning to develop extended reasoning behavior and made downloadable weights available under permissive terms. DeepSeek said results were broadly comparable with OpenAI o1 on several math, coding, and reasoning benchmarks, but the comparison should be read task by task rather than as universal superiority.
Assessment
Why it matters
R1 made frontier-style reasoning inspectable, downloadable, and distillable, accelerating open research and changing the pace of model competition.