DeepSeek is a Chinese AI lab whose open-weight models achieved near-frontier performance at dramatically lower training cost. Its R1 reasoning model was a genuine inflection point for the open-model ecosystem.
DeepSeek is an AI research company based in China that releases its models with open weights — meaning anyone can download them, run them on their own hardware and build products on them without paying per use.
It became widely known when its reasoning model, R1, performed comparably to the leading commercial reasoning models on maths and coding benchmarks, while being free to download.
Two reasons. The performance was much closer to the expensive frontier models than anyone expected from an open release. And the reported training cost was far lower than the sums usually associated with models of that capability.
That combination challenged an assumption the industry had been operating on — that frontier capability required enormous capital and would therefore stay concentrated in a handful of well-funded labs.
If you want a capable model you can run yourself, without per-token costs and without sending data to a third party, DeepSeek models are among the strongest options available.
If you want the easiest possible experience, a hosted assistant like ChatGPT or Claude is simpler. DeepSeek is most valuable to people who specifically want self-hosting or very low API cost.
DeepSeek-V3 is a general-purpose model using a mixture-of-experts architecture, where only a fraction of total parameters activate for any given token. This is what allows a very large model to run at the inference cost of a much smaller one.
DeepSeek-R1 is the reasoning model — it generates an extended chain of thought before answering, in the same family as other test-time-compute approaches. It is strongest on mathematics, competitive programming and multi-step logical problems.
Distilled versions of R1 exist at smaller sizes, transferring much of the reasoning capability into models that run on ordinary hardware.
Self-hosted — download weights from Hugging Face and serve with vLLM, Ollama or similar. Full data control, no per-token cost, your own infrastructure to run.
DeepSeek API — hosted by DeepSeek at notably low prices. Fastest route, but requests go to servers in China, which is a material consideration for many organisations.
Third-party hosting — several inference providers serve DeepSeek models from Western data centres, which resolves the data residency question while keeping the cost advantage.
Against Llama and Qwen, DeepSeek is generally stronger on reasoning and coding and comparable on general tasks. Against closed frontier models, it trails on the hardest problems and long-context work while being adequate for a wide range of practical tasks.
The decision is rarely about raw capability alone — it is about whether open weights, self-hosting and cost matter more than the last increment of performance.
The mixture-of-experts design routes each token through a small subset of specialised sub-networks. Total parameter count is large; activated parameters per token are far smaller. This decouples model capacity from inference cost and is a significant part of why the economics work.
DeepSeek also published details of training efficiency techniques that were widely studied afterward, which is itself unusual — most labs at that capability level publish very little.
Using the DeepSeek-hosted API means sending data to infrastructure in China, subject to Chinese law. For many organisations this is disqualifying on its own, regardless of the model's quality.
Self-hosting removes the issue entirely — the weights are just a file, and running them locally involves no connection to DeepSeek. This distinction is frequently collapsed in public discussion and is worth keeping clear.
Models trained in different regulatory environments reflect them. DeepSeek models decline or deflect on topics sensitive in China in ways Western models do not, and behave differently on some alignment dimensions.
For most technical and commercial applications this is irrelevant. For anything touching on politics, history or journalism, evaluate the behaviour against your requirements rather than assuming equivalence.
Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.
Nothing is sent from here. The link carries only this page’s title and address.