# DeepSeek

> A Chinese AI lab whose open-weight models showed that top AI could be built for less.

- Level: Basic
- Topics: Tech companies, AI
- URL: https://tidbit.legios.com.ar/deepseek/
- In Spanish: [DeepSeek](https://tidbit.legios.com.ar/es/deepseek/)

**DeepSeek** is an AI company based in Hangzhou, China, founded in 2023 by Liang Wenfeng and funded by High-Flyer, the quantitative investment fund he co-founded. It builds large language models, offers them in a free chat app and sells access to them through an API at low prices.

It drew worldwide attention in January 2025 with **DeepSeek-R1**, a reasoning model that thinks step by step before answering, built on top of its earlier model DeepSeek-V3. R1 performed close to the best models from American labs on many tests, but DeepSeek published its **weights** under a permissive license, so anyone could download it and run it, together with technical reports explaining how it was trained. The reports also described tricks to train efficiently, such as a mixture of experts design and using reinforcement learning to teach reasoning, and a training cost far lower than people expected, although that figure covered only the final training run.

Its arrival mattered for several reasons. It suggested that strong models did not require the biggest budgets, it came from China despite US limits on exports of the most advanced AI chips, and it gave researchers everywhere an open model to study and adapt. It also raised concerns: the app stores data on servers in China, some governments banned it on official devices, and its hosted service avoids or follows official lines on topics that are sensitive in China.

## An analogy

A small kitchen that cooks a dish as good as the famous restaurants for a fraction of the cost, and then publishes the full recipe for anyone to try at home.

## Keep in mind

- Chinese AI company founded in 2023 in Hangzhou, backed by the fund High-Flyer.
- DeepSeek-R1 (2025) rivaled top models and was released with open weights.
- Its technical reports showed how to train strong models more efficiently.
- Its app has raised privacy and censorship concerns in several countries.

## Go deeper

- [DeepSeek on GitHub (models and reports)](https://github.com/deepseek-ai)
- [DeepSeek on Wikipedia](https://en.wikipedia.org/wiki/DeepSeek)
- [DeepSeek official site](https://www.deepseek.com/)

## Related terms

- [Artificial intelligence](https://tidbit.legios.com.ar/artificial-intelligence/): Getting computers to do things that, until recently, only people could do.
- [API](https://tidbit.legios.com.ar/api/): The menu of things one program lets another program ask for.
- [Model parameters](https://tidbit.legios.com.ar/model-parameters/): The billions of numbers a model learns in training; '70B' means 70 billion of them.
- [Reinforcement learning](https://tidbit.legios.com.ar/reinforcement-learning/): Learning by trial and error, guided by rewards for good outcomes.
- [Training and inference](https://tidbit.legios.com.ar/training-and-inference/): Training builds the model once; inference is using it, millions of times a day.
- [Chip](https://tidbit.legios.com.ar/semiconductor-chip/): A sliver of silicon carrying billions of tiny switches that compute and remember.
- [Server](https://tidbit.legios.com.ar/server/): A computer whose job is to answer requests from other computers, all day, every day.
- [Open-weights models](https://tidbit.legios.com.ar/open-weights-models/): AI models whose trained numbers anyone can download and run on their own machines.