Skip to content
BENIBIX Learn

Llama, Mistral, DeepSeek, Qwen

Level 1 · The AI labs

The four best-known open-weight families and the labs behind them. What each one shares, and what its license allows you to do.

Course contentsLesson 3 of 13

After this lesson you can

  • Name the labs behind four open-weight families
  • Say what makes DeepSeek-R1 stand out
  • Explain why the open file is not the app

11 mincovers 5 glossary entries

Worth reading first: AI Labs and Model Families

Who makes the best-known open models

The best-known open-weight models come from four labs. Meta released Llama in February 2023. Mistral AI, founded in Paris in April 2023, came soon after. DeepSeek, in China, shook the field in January 2025 with DeepSeek-R1. Alibaba's Qwen team makes Qwen. You can download models from all four and run them yourself.

Meta, the social media firm, makes Llama through its AI arm, Meta AI. Llama helped start the open-weight boom, and thousands of tuned versions were built on it. Its license allows most uses. But under the Llama 2 license, a firm with over 700 million users a month had to ask Meta first.

Mistral AI is the best-known lab in Europe. Its founders came from Google DeepMind and Meta. It shares some models with open weights and sells others only through its API. It also runs its own chat app. Many European firms pick it so that their data stays under European rules.

DeepSeek was founded in Hangzhou in 2023 and is owned by a Chinese hedge fund. In January 2025 it released DeepSeek-R1, a model that thinks before it answers. It matched top closed models on many maths and code tests. Its weights came under the MIT license, which lets almost anyone use them, even in a paid product.

Qwen comes from Alibaba's Qwen team, part of a big Chinese firm. It comes in many sizes, from small models for a phone to very large ones. Many Qwen models use the Apache 2.0 license, which allows business use, though not every one does. So many people build their own models on top of Qwen.

The weights and the app are two separate things. DeepSeek's own app sends what you type to DeepSeek's servers. The same model, downloaded and run on your own server, sends nothing anywhere. The open file is what gives you that choice.

People call all four open source, and they are something less. Most share the trained file but not the data it learned from or the full recipe. Each comes with its own license. Read the license of the exact model you pick, because two models from one lab can have different terms.

An open-weight model leaves the lab as a file. Where it runs after that is up to you.

Worked example

Lucy's bank keeps its data in-house

Lucy's bank may not send customer letters to any outside firm. So a closed model reached over the internet is out.

She downloads a small Qwen model and checks its license. It is Apache 2.0, so she runs it on one of the bank's own servers.

The words for this lesson5

Check yourself

Can you run DeepSeek and keep your data away from it?

Finished this lesson?

Marking it complete also marks its 5 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.