DeepSeek is a Chinese AI lab backed by High-Flyer Capital that shocked the industry in early 2025 when their R1 reasoning model matched OpenAI's o1 at a fraction of the training cost. Their V3 general-purpose model and R1 reasoning model are both available open-weight, and their API is among the cheapest frontier-class options in the world.
DeepSeek publishes open weights for V3 (general) and R1 (reasoning) on Hugging Face, under a permissive license. Self-hosting is identical to Llama — vLLM, llama.cpp, etc. Their commercial API offers the same models at a very low per-token cost, though it routes through Chinese infrastructure. For sensitive data, self-host the weights on your own infrastructure.
Free to self-host. DeepSeek commercial API: V3 at $0.27 input / $1.10 output per M tokens. R1 reasoning: $0.55/$2.19. Among the cheapest frontier pricing available. Off-peak discounts further reduce costs by up to 75%.
Broadly deployed in the open-source community. Popular with budget-conscious builders and researchers, especially for reasoning-heavy tasks. Avoided by some enterprises due to China data residency concerns.