DeepSeek R1

DeepSeek R1 was DeepSeek's first open-source reasoning model, released in January 2025. It popularised visible chain-of-thought reasoning and rivalled OpenAI's o1 at a fraction of the cost. It is no longer DeepSeek's current model: reasoning now runs inside DeepSeek V4, and the deepseek-reasoner API alias that served R1-style output was retired on 24 July 2026.

The R1 weights are still available on Hugging Face under an MIT licence, so you can download and self-host them for free. What is gone is the hosted route: the deepseek-reasoner API alias was retired on 24 July 2026, so new API work should target deepseek-v4-flash or deepseek-v4-pro.

Only for research, reproducibility or offline self-hosting. For anything production-facing, DeepSeek V4 supersedes R1 on reasoning, coding and cost, and it is the only line DeepSeek still serves through its API.

Yes. The R1 weights remain downloadable from Hugging Face and run under vLLM, SGLang or Ollama (distilled variants for smaller GPUs). Self-hosting is now the main reason to still pick R1 over V4.

R1 was built for step-by-step reasoning: mathematics, logical deduction, code generation and multi-step analysis, with its chain of thought exposed in the output. Those capabilities now live in DeepSeek V4, which reasons internally and scores higher on the same benchmarks.

R1 shipped with content filtering, and because the weights are open they can be audited and run entirely on your own infrastructure — which is the strongest privacy option available. Note that R1 no longer receives updates or safety patches from DeepSeek.