From the archive
DeepSeek releases R1 reasoning models and API access
R1 becomes available through chat, a hosted reasoning endpoint, and downloadable weights, with smaller distilled variants.
DeepSeek released R1, a model focused on reasoning tasks such as mathematics and code. The release included a hosted chat option, application programming interface (API) access, model weights, and six smaller distilled models.
Hosted or self-managed reasoning
The hosted endpoint uses the model name deepseek-reasoner. At launch, input pricing distinguished cached tokens from uncached tokens, with output billed separately. That makes repeated long inputs and the amount of generated reasoning relevant to the total cost of a request.
The main R1 release uses the MIT license, allowing experimentation and commercial use under its terms. The smaller distilled models offer another route for running reasoning workloads on different hardware capacities.
Access beyond a chat page
The launch gives applications a choice between sending requests to DeepSeek and managing an inference deployment. Those routes involve different hardware, operating, and data-handling requirements. The DeepSeek profile covers its hosted models and billing structure.