DeepSeek-V3 and R1

DeepSeek-V3 and R1 are downloadable language models for developers who want to run text generation and reasoning on their own hardware. V3 is a general language model; R1 focuses on problems that require extended reasoning.

• Text generation V3 provides base and chat models built with a mixture-of-experts architecture, with capabilities in language, math and coding.
• Reasoning R1 works through math, code and other reasoning tasks, and builds on DeepSeek-V3-Base.
• Smaller models Distilled R1 models use Qwen2.5 and Llama 3 bases, offering smaller alternatives to the full R1 model trained with examples generated by R1.
• Reasoning research R1-Zero lets researchers examine reasoning learned through reinforcement learning without preliminary supervised fine-tuning, though its answers can repeat, mix languages and be difficult to read.

Both families have weights available through Hugging Face for local use. DeepSeek also provides R1 through a hosted chat website and an OpenAI-compatible API; those services run remotely rather than on your hardware. Hugging Face Transformers doesn't directly support the full R1 models.

The V3 and R1 code repositories use the MIT License. V3's base and chat model weights have a separate Model License that permits commercial use.