What Is DeepSeek? Hosted Models and Open Weights
DeepSeek includes hosted services and several model series with published weights. This article explains API aliases, thinking mode and self-hosted versions, so that old tutorial names are not mistaken for the current underlying model.
Meet this model
DeepSeek is a generative AI company and model brand, and the name is also commonly used for its chat app. It offers hosted APIs and has released model weights for self-hosting. A hosted API name may be an alias pointing to an updated model, while downloaded weights correspond to a particular release. This distinction helps determine whether old code still calls the original model and whether a hosted experience can be reproduced locally.
Current API aliases differ from historical releases
On October 20, 2026, the official quickstart specifies deepseek-flash, served by DeepSeek-V4.1-Flash-1017. The older name deepseek-v4-flash is still accepted, but its original model has been retired and requests are served by the newer Flash model. The underlying model can therefore change even when the calling code does not.
The API offers OpenAI-compatible and Anthropic-compatible interfaces, reducing client changes without guaranteeing identical parameters or tool behavior. When migrating an older project, confirm the actual model name and thinking settings in addition to the address and credentials. A similar request format does not replace comparisons of responses and tool requests.
Thinking mode is not the same as online verification
Thinking mode allows additional reasoning before an answer; it does not turn training knowledge into live information. A question about a software release announced today still needs search or the current release page as input. Showing reasoning_content does not prove that a web page was visited or that a factual claim is supported by an original source.
For example, provide the same constraints when comparing two technical proposals, ask which source passage supports each requirement, and check omitted conditions. To evaluate thinking mode, keep inputs identical and distinguish reasoning content from the final answer. This article does not measure DeepSeek’s reasoning accuracy or treat provider benchmark scores as your task success rate.
Check the exact release before downloading weights
Open weights offer a separate way to use the models. The official DeepSeek-V3.2-Exp repository identifies it as an experimental release using DeepSeek Sparse Attention to explore long-context efficiency and provides local inference instructions. It should not be described as the same weights currently serving hosted Flash: the release and inference environment differ.
The V3.2-Exp repository and weights use the MIT license; keep the license for the downloaded release. Running the full model also requires a compatible inference framework and hardware, and a series number does not specify memory requirements. For distilled or quantized releases, inspect the exact model card rather than transferring the full model’s capability claims to a smaller version.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.