THE PATTERN
● Qwen 3.8-Max: custom licence, not Apache. Text-only.
● Kimi K3: modified MIT. The modification is the part to read.
● GLM-5.3: weights held two weeks. Flagship licence still unstated. Flash was MIT.
● The OSI position: traditional software licensing does not automatically secure the freedoms an AI system requires.
Downloadable is not the same as open
For two years the shorthand held well enough. If you could download the weights, you could do roughly what you wanted with them, and the licence was usually Apache 2.0 or MIT.
That has quietly stopped being true, and the change happened without much comment because the headline word — open — did not change.
| Release |
Terms |
What to check |
| Qwen 3.8-Max |
Custom licence |
What the custom terms restrict that Apache would not |
| Kimi K3 |
Modified MIT |
The modification. MIT semantics do not carry over automatically |
| GLM-5.3 flagship |
Not stated |
Whether terms are published before you commit |
| GLM-5.3-Flash |
MIT |
Genuinely permissive. The variant, not the flagship |
WHY SOFTWARE LICENCES DO NOT TRANSFER CLEANLY
MIT and Apache were written for source code. You can read code, modify it, and know what you have.
A weights file is not readable in that sense. You cannot inspect what is in it, reproduce it without the training data, or verify what it was trained on. A licence granting you freedom to modify something you cannot inspect grants less than the same words grant for source.
That is the substance of the Open Source Initiative's warning, and it is a fair point rather than a definitional quibble. Applying a code licence to weights uses familiar words for an unfamiliar object.
What actually matters commercially
Most teams reading a model licence care about four things, and none of them are settled by the word open.
- Commercial use. Permitted at all, and at what scale? Some licences add revenue or user thresholds.
- Output ownership. Who owns what the model generates, and are there restrictions on what you may do with it?
- Derivative work. May you fine-tune and redistribute, and under what terms?
- Term stability. Can the licence change for future versions in a way that strands your deployment?
The last one is the one people skip. A permissive licence on version 3 does not oblige anyone to be permissive on version 4, and if your pipeline assumes upgrades, that is a real exposure.
The other half of the problem
Licensing is not the only reason downloadable and available are drifting apart. Z.ai held GLM-5.3 weights for two weeks citing cyber capability that emerged faster than expected. OpenAI gated a cyber-capable model in the same fortnight. Anthropic disclosed a frontier model it will not ship.
So there are now two independent reasons a model you expect to download might not arrive: the terms may not permit what you need, and the weights may not be published at all. Both are new, and both landed within about six weeks.
What to do before you build on any of them
| Check |
Why |
| Read the licence file, not the announcement |
Announcements say open. Files say what you may do |
| Search the text for modified or custom |
Those words mean the standard semantics do not apply |
| Keep a copy of the terms you agreed to |
Terms change. What applied at download is what you relied on |
| Assume the next version may differ |
Nothing obliges a lab to keep terms stable across versions |
| Keep a fallback configured |
Availability is now a variable, not a constant |
None of this is legal advice, and anything commercially significant belongs with a lawyer rather than a checklist.
Sources
FAQ
Are open-weight models still open source?
Not always in the sense the term implies. Recent releases have shipped under custom and modified licences rather than Apache or MIT, and the Open Source Initiative argues that traditional software licensing does not automatically secure the freedoms an AI system requires.
What is openwashing?
Presenting a model as open when the licence restricts what open normally implies — commercial use, redistribution, or derivative work.
Which recent models are genuinely permissive?
GLM-5.3-Flash shipped under MIT. The GLM-5.3 flagship licence has not been stated, Kimi K3 is modified MIT, and Qwen 3.8-Max used a custom licence.
What should I check in a model licence?
Whether commercial use is permitted and at what scale, who owns the output, whether you may fine-tune and redistribute, and whether the terms could change for future versions.
Can a licence change after I have deployed?
Terms you agreed to at download generally govern that copy, but nothing obliges a lab to keep them stable for future versions. Keep a copy of what you agreed to.
Is this legal advice?
No. Anything commercially significant belongs with a lawyer.