DeepSeek-V3 and R1

DeepSeek-V3 and DeepSeek-R1 are downloadable language models for developers who want to run text generation and reasoning on their own hardware. V3 is a mixture-of-experts text model. R1 builds on DeepSeek-V3-Base and focuses on reasoning tasks such as math and coding. The V3 and R1 repositories describe their respective models; the R1 repository links to downloadable weights.

R1 is trained to work through complex problems and check its reasoning. DeepSeek also publishes R1-Zero, a research model trained with reinforcement learning alone. R1-Zero can repeat itself, mix languages and produce harder-to-read answers; R1 uses additional training to address those issues. Smaller R1 models distilled from the original are available on Qwen2.5 and Llama bases, giving people more choices for local use than the full-size model.

These model families are distinct from DeepSeek’s hosted chat and API services. A local deployment uses downloaded weights and a compatible inference setup. The model repositories and cards are the places to check the terms and requirements for the particular checkpoint you choose.