Open Source Series · What Open Source Actually Opens

Real vs. Fake Open Source: How to Read a License

Almost every vendor says its models are "open source," but the word means very different things from one to the next. Some let you build a business on them with no conditions, some require a separate contract once your user base grows, and some have only published a paper. Learning to judge for yourself beats memorizing who is open.

Three questions, and you have your answer

You do not need to read the license end to end. Ask the following three questions in order and you will have placed almost any model on the openness spectrum.

Question 1Can you download the weight file?

Everything else depends on this. If a model is reachable only through an API, then whatever the marketing says, you never get the control described in the last section. This one question rules out more than half the field.

Question 2Can you use it commercially, and are there conditions attached?

This is where people get caught. Some licenses permit commercial use but cap your user base; others limit you to research. Skip this check before you build a product and you will be scrambling for a license right when your numbers start climbing.

Question 3Can you train new models on its output?

In other words, can you distill from it. This decides whether you can build a model of your own on top of it, which is what the next two sections unpack. Vendors disagree on this point more than any other.

Reading is not the same as knowing how to use it, so take a model you are actually considering and walk it through. The wizard below is the same three questions, one at a time, and it tells you which tier the model lands in.

Can you download the weight file?
Question 1 of 3
Five vendors, one yardstick

Each vendor below gets the same four questions. The fourth one — whether the training data is public — is included to make one thing visible: in the world of large models, the word "open source" has already been redefined.

Qwen3, all sizes
No extra conditions
Apache 2.0
  • YWeights downloadable
  • YCommercial use, no user cap
  • YCan be used to train new models
  • NTraining data not disclosed
Mistral 7B
No extra conditions
Apache 2.0
  • YWeights downloadable
  • YCommercial use, no user cap
  • YCan be used to train new models
  • NTraining data not disclosed
DeepSeek-R1
No extra conditions
MIT License
  • YWeights downloadable
  • YCommercial use, no application needed
  • YDistillation explicitly permitted
  • NTraining data not disclosed
Llama 3
Conditions on commercial use
Llama 3 Community License
  • YWeights downloadable
  • LSeparate license required above 700M monthly actives
  • YCan be used to train new models
  • NTraining data not disclosed
GPT-4 / Claude / Gemini Ultra
API only
No public weight license
  • NWeights not downloadable
  • NNo local deployment
  • NCannot be used for training
  • NTraining data not disclosed
Based on the vendors' own license texts: Meta Llama 3 Community License, clause 2 (more than 700 million monthly active users in the preceding calendar month requires a separate commercial license from Meta); Mistral's official release notes (Apache 2.0, without restrictions); DeepSeek's official announcement of 2025-01-20 (MIT License, distillation to train other models explicitly permitted); Qwen3 official release notes. Verified 2026-08-07.
Sorted into three tiers
Tier 1 · Open weights with no extra conditions
Covered by standard open-source licenses such as Apache 2.0 or MIT. Downloading, modifying, commercial use, redistribution, and training new models all happen without asking permission, and there is no user-count threshold. This is the easiest case for a legal review.
Examples: Qwen3 across all sizes, Mistral 7B, DeepSeek-R1
Tier 2 · Open weights with custom terms attached
A license the vendor wrote itself, usually adding limits on commercial scale, use cases, or attribution. Technically it is fully usable; commercially you need to work out in advance whether you will hit the line.
Example: the Llama 3 family
Tier 3 · API only
No weights, which also means no local deployment, no on-premise option, and no further training. The capability may well be the strongest available, but what you rent is the right to use it.
Examples: GPT-4, Claude, Gemini Ultra
One fact that has to be stated plainly
None of the five vendors above has released its training data. By the standards of traditional software, the equivalent of source code would be the training data plus the training code, and shipping only the compiled artifact does not count as open source. What the large-model field calls open source actually means open weights. It is a watered-down definition the industry has tacitly accepted, and you should keep its boundaries in mind whenever you use the word.

It is also worth noting that licenses change. In the 2.5 generation Qwen used a tiered license: most sizes were Apache 2.0, while 3B and 72B had separate terms. Only with Qwen3 did the whole family standardize on Apache 2.0. What matters during selection is the license of the specific version you plan to use, not your general impression of the company.

Most sizes in the Qwen2.5 family use Apache 2.0; the 3B model uses the Qwen Research license and the 72B model uses the Qwen license. All eight models in the Qwen3 family use Apache 2.0. Source: Qwen official release notes.
Take this checklist with you

Next time you evaluate a model, walk through it in order:

1. Go to the official repository and read the license file itself, not the press release or the landing page.
2. Confirm the license for the specific size you intend to use; it can differ within one family.
3. If it is a custom license, look specifically for limits on user scale, use case, and geography.
4. Confirm whether training is permitted, since that determines whether fine-tuning and distillation stay open to you later.

The next section takes a different angle: why these vendors open-source at all. Open-sourcing costs money, released weights can never be taken back, and none of them is doing charity.