What "open-source AI" means after 2025
In 2025, more than one popular model was called open while its license said otherwise. If you run a business or a school anywhere, that gap is not a detail. It decides who can see your data, what you pay, and whether you can keep the system running without asking permission.
TL;DR Open weight is not open source. Many models marketed as open carry licenses with usage limits and no-competing-model clauses. For any organisation that wants to self-host, the license decides your data sovereignty, your running cost, and your uptime. Before you deploy anything, read the license text itself, not the blog post about it.
Open weight is not open source
The phrase open-source AI is used for at least three different things. The first is a model you can download and run. The second is a model whose weights are public but whose license restricts how you use them. The third is a model with both public weights and a permissive license such as Apache-2.0 or MIT, where you can modify, host, and ship it commercially with almost no gate. Only the third matches what open source has meant in software for thirty years.
Most models people call open in 2025 and 2026 sit in the second group. The weights are out. The license is not. That is a real and useful thing. It is just not the same as open source, and the difference shows up the moment you try to build something serious on top of it.
What changed in 2025
Through 2025, several widely used models kept their weights public but tightened the rules around them. Some added acceptable-use policies that forbid certain applications. Some blocked training a competing model above a parameter threshold. Some limited use by organisations above a revenue or funding size. None of this makes the model useless. All of it changes what you are allowed to do with it.
The shift was not one dramatic event. It was a slow move from open-weight-with-strings toward open-weight-with-more-strings. For a team planning a year of work on top of a model, that drift is the risk. The model you can use freely today may carry new limits by the time you ship.
Why this matters
- A cloud API holds your prompts and your data on someone else’s server, often in another country. A self-hosted open model keeps that data inside your building.
- An API bill scales with usage. A self-hosted model has a known, one-time hardware cost and then runs.
- When the network drops, a cloud feature stops. An offline model keeps working.
- Some licenses stop you from serving certain sectors or from building a competing product. If your plan touches those areas, the license decides your roadmap.
For a school in a town with poor connectivity, or a small shop with thin margins, these are not theoretical points. They are the difference between a tool that works in March and one that quietly stops.
How to read a model license
- Is the weight license Apache-2.0, MIT, or similar permissive? If yes, you can usually self-host, modify, and ship commercially.
- Is it a community or open-weight license? Read the acceptable-use policy line by line.
- Does it restrict commercial use, output types, or geographic regions? Write those down before you build.
- Does it claim rights over models you train on top of it? If so, your own work may not be yours to keep.
What we run, and what it costs
As a worked example rather than a claim about any one vendor, a 7-billion-parameter model runs on a single consumer graphics card with enough memory, and a 14-billion-parameter model runs on a single workstation-class card. The exact speed depends on the hardware, the quantization, and the date you run it. When a benchmark is published, the honest ones state the machine, the quantization level, the date, and who ran it. A number without those four facts is not a benchmark. (P2: independent and vendor runs; always verify on your own hardware before you buy.)
Common mistakes we see
- Trusting the homepage label and skipping the LICENSE file.
- Assuming open weight means free for commercial use.
- Building on a model whose license forbids your sector, then discovering it at procurement.
- Quoting a benchmark that omits the hardware and the date.
Each of these is avoidable. None of them shows up in a demo. They show up later, in a contract or an audit, which is the worst time to learn them.
The license is the part of the model you actually sign. Everything else is a brochure.
What to do this week
If you lead a team that is about to adopt a model, set a standing rule. No model enters a project until someone has read its license and written one line on what it permits. Keep that line in the project file. It takes ten minutes and it prevents the expensive surprises. This is the unglamorous part of doing AI seriously, and it is where most organisations cut corners.
This is written not to scare you away from open models. Quite the opposite. Open-weight and open-source models are the reason a school or a small shop anywhere can run capable AI without sending every question abroad. The point is to know exactly what you have downloaded, so you can stand on it instead of borrowing it.
Where this series goes
This is the first post in a short series on owning your AI. Next: a full offline deployment inside a school, with the commands and the hardware, then a briefing on what state governments should ask before they sign a cloud contract. Each post follows the same rule. Show the license, the hardware, and the number, or do not make the claim.