January 2025 was the DeepSeek wake-up call. An open-weights model matched the frontier and cost a fraction per token. The market lost a trillion dollars in a week.
DeepSeek V3: the flagship — 671B total parameters, 37B active (MoE), frontier-ish performance on reasoning, code and math. Open weights, MIT license.
DeepSeek R1: the reasoning model. Trained with reinforcement learning rather than human-written reasoning traces. Cheap, open, and shockingly good at competition math and code.
Licensing: MIT for the weights, which means commercial use allowed, fork allowed, fine-tune allowed.
Access: their own API is very cheap (~$0.27/1M input, $1.10/1M output). Available on Together, Fireworks, OpenRouter, Groq.
The caveats: censorship around Chinese politics topics, and API reliability has occasional hiccups during peak traffic.
The strategic impact: DeepSeek proved you don't need $100B to train a frontier model. It forced every lab to justify their spend. Prices dropped across the board. Users won.
For most teams, DeepSeek V3 is the new default for cost-sensitive inference. Test it on your workload — you'll be surprised.