TL;DR: DeepSeek is a Chinese AI lab that releases open-weight models, best known for strong reasoning ability delivered at a fraction of the training cost the field assumed was required. Its releases are downloadable under permissive licenses, and you can run them yourself, use its first-party API, or reach them through third-party inference providers.
What it is and how you use it
DeepSeek is based in Hangzhou and grew out of High-Flyer, a Chinese quantitative hedge fund whose GPU clusters and engineering culture it inherited. That origin shows in the work: the lab is known less for product polish than for aggressive efficiency research - training techniques and mixture-of-experts architectures that squeeze frontier-level capability out of far less compute than its Western peers spend.
The family's defining moment came in early 2025, when its open-weight reasoning model matched leading closed models on hard math and coding problems while its reported training budget undercut theirs by an order of magnitude. The release reset the industry's assumptions about how much money frontier capability requires and proved that open weights and top-tier reasoning were not mutually exclusive. Since then, the durable identity has held steady: open releases, strong reasoning and coding, unusually low cost - both to train and, through its API, to use.
Access comes in three forms. The weights themselves are on Hugging Face under permissive licenses (MIT for key releases), so you can self-host on your own GPUs or run distilled variants as a local LLM. DeepSeek's own API is priced aggressively but is hosted in China, which many organizations exclude on data-governance grounds. The middle path is common in practice: inference providers and Western clouds host the open weights on their own infrastructure, so you get the models without sending data to the lab that made them.
Where it typically fits: reasoning-heavy and coding workloads where budget matters, research and experimentation that benefits from inspectable open releases, and self-hosted deployments that want near-frontier capability without a per-token contract. The counterweights are governance questions around the first-party API's jurisdiction and the operational load of serving large models yourself - which is exactly why the third-party hosting path is so widely used.
Where it sits in the AI stack
DeepSeek sits at the model layer as an open family: the weights are the product, and the serving arrangement - self-hosted, first-party, or third-party - is your decision:
Key tools and implementations
-
Open-weight downloads
Full models and smaller distilled variants on Hugging Face under permissive licenses.
-
DeepSeek API
The first-party, OpenAI-compatible endpoint - very low cost, hosted in China.
-
Third-party providers
Western inference providers and clouds serving the same open weights on their own infrastructure.
-
Local and serving runtimes
Ollama runs distilled variants on workstations; vLLM serves the full models on self-managed GPUs.
Related entries
- Qwen Alibaba's family of open-weight models spanning many sizes, strong in multilingual and coding tasks.
- Reasoning models Models trained to work through problems step by step before answering, trading time and tokens for accuracy.
- Open-weights models Models whose trained weights are published for anyone to download, run locally, and fine-tune.
- Mixture of experts An architecture that activates only a few specialist subnetworks per token, so a huge model runs at a smaller model's cost.