DeepSeek-R1

DeepSeek built this reasoning model by running extensive reinforcement learning over its DeepSeek-V3 base, so it writes a long thinking trace inside think tags before the final answer, which pays off on competition math, coding and logic problems. The flagship is a mixture-of-experts network with 671 billion total parameters, about 37 billion active per token, and a 128K context window; an experimental R1-Zero sibling skipped supervised tuning entirely. A May 2025 update, R1-0528, deepened reasoning, cut hallucinations, improved function calling and added system prompt support. Serving the full model takes a multi-GPU server with several hundred gigabytes of accelerator memory through vLLM, SGLang or a similar engine, while six distilled checkpoints built on Qwen 2.5 and Llama 3 scale down to 1.5B parameters, small enough for a single consumer GPU. Weights and code are MIT licensed with commercial use and distillation allowed, though the Llama-based distills also carry Meta's license terms. DeepSeek suggests a temperature near 0.6. A strong pick for anyone wanting an openly licensed reasoning model to self-host or study.