DeepSeek made global headlines in early 2025 when its R1 reasoning model matched or approached the performance of much more expensive closed models at a fraction of the training cost, triggering a broad reassessment of how much compute state-of-the-art AI actually required. The lab has kept releasing open-weight models since then: V3 for general-purpose tasks, R1 for step-by-step reasoning, and a dedicated Coder line tuned specifically for programming tasks.
All of it runs on a mixture-of-experts architecture with a 128K context window, and crucially, the code and weights are released under the MIT license, meaning anyone can download, modify, fine-tune, and deploy the models commercially without licensing fees or usage restrictions. That openness, combined with genuinely competitive benchmark performance, made DeepSeek one of the few labs outside the US with models developers actually reach for by choice rather than necessity.
It's completely free to self-host, with the obvious trade-off that running a model at this scale takes real infrastructure, or developers use one of the many third-party providers now hosting DeepSeek's models as an API.