Names, versions and owners change fast. Checked October 2026.
In plain words
DeepSeek is three things with one name: a company in Hangzhou, China; a chatbot app; and a family of large language models that power it. The company was founded in 2023 and is backed by High-Flyer, a Chinese investment fund.
It became world news in January 2025. Its model R1, a reasoning model that works through problems step by step, performed close to the best American models on maths and coding tests. Its app briefly became the most downloaded free app in the US App Store.
What sets DeepSeek apart is that it publishes its models with open weights under a permissive licence, so anyone can download, run and change them. It releases new versions often; the box above shows the newest.
Everyday example
A hospital with many specialists
A big hospital has dozens of specialists, but a patient with a sprained ankle sees only the orthopaedic doctor and a nurse, not every doctor in the building. DeepSeek's large models work like that. They contain many small expert parts, and a router sends each piece of text only to the few experts that suit it. The hospital is huge; each visit is small.
Where the example breaks: Hospital specialists are chosen by topic. DeepSeek's experts are not neatly "the maths expert" or "the grammar expert"; the model learns its own split, which is often hard for people to interpret.
Latestas of October 2026
- Newest models
- DeepSeek V4.1 Flash (10 September 2026, updated 17 October 2026) and V4 Pro (previewed 24 April 2026, full release 13 August 2026)
- Where to use it
- DeepSeek's website, app and API; weights on Hugging Face and ModelScope
- Licence
- MIT licence for the model weights
These details change often. The rest of this page explains what it is, which stays true.
How it actually works
Two technical choices explain much of DeepSeek's attention: how its models are built, and how R1 learned to reason.
Mixture of experts
DeepSeek's large models are a mixture of experts. Its V4 Pro model, from 2026, holds about 1.6 trillion parameters, but only around 49 billion are used for each token. That keeps each answer cheaper to compute than the total size suggests.
Reasoning learned by trial
For R1, DeepSeek used large-scale reinforcement learning: the model tried problems with checkable answers, such as maths, and was rewarded when it got them right. It learned to write out long working steps before answering.
Open weights, MIT licence
R1, later versions of V3, and the V4 models were released under the MIT licence, which allows commercial use and changes. Many companies run DeepSeek models on their own servers, and smaller distilled versions run on a single computer.
Long context
Its 2026 V4 models accept up to about a million tokens of input in one go, a very large context window.
- Company
- Hangzhou DeepSeek, founded July 2023 in China; CEO Liang Wenfeng
- Breakthrough model
- DeepSeek-R1, released 20 January 2025
- Known for
- Open-weight models and low-cost mixture-of-experts designs
- Licence
- Model weights under the MIT licence; the apps themselves are proprietary
Questions people ask
Is DeepSeek the same as ChatGPT?
Is DeepSeek open source?
Its model weights are public under the MIT licence, so you can download and modify them. It has not released all of its training data, so many people call it open weights rather than fully open source.
Is DeepSeek safe to use?
Using the app means your messages go to DeepSeek's servers in China, and some governments have restricted it on official devices over data concerns. Running the open model on your own computer keeps data local. As with any chatbot, avoid sharing sensitive information.
Why did DeepSeek shake markets in January 2025?
R1 matched leading American models on several maths and coding tests, yet anyone could download it. Investors began to question how many expensive chips AI would really need, and shares of chipmakers such as Nvidia fell sharply on 27 January 2025.