Open-Weight Models: What They Mean and How to Choose
Read any announcement about a new downloadable AI model and you will hit the phrase "open-weight models" in the first paragraph. Nobody ever stops to define it, and it is not the same thing as open source — a distinction that matters a great deal the moment you want to use one of these models in something real. So let me define the term properly, walk through the families you will actually run into, show you which license clauses to read before you download, and give you a short way to choose.
No setup instructions here. This is the map I’d want before starting on multi-gigabyte downloads.
What "open weight" actually means
A trained AI model is, at heart, a very large pile of numbers. Those numbers are the weights (also called parameters): the values the model learned during training, which determine how it responds to anything you type.
A model is open-weight when the people who trained it publish those numbers for anyone to download. You get the finished brain. You can run it on your own machine, inspect it, fine-tune it, and keep using it after the company that made it loses interest.
What you usually do not get is everything else:
- The training data — what it read, and in what proportions.
- The training code and recipe — the exact process that produced those numbers.
- A detailed log of the fine-tuning and safety work layered on top.
So an open-weight model is closer to a compiled program than to source code. You can run it and modify it, but you cannot rebuild it from scratch or fully audit how it came to think the way it does.
Open weights vs. open source
This is where most of the confusion lives, and it is not pedantry — it decides what you are legally allowed to do.
"Open source" is a term with a definition behind it. For software, the Open Source Initiative has maintained that definition for decades: free use for any purpose, no discrimination against fields of use, and the right to redistribute modifications. In 2024 the OSI published an equivalent definition for AI systems, and it asks for three things together: the parameters, the code used to train and run the model, and enough information about the training data that a skilled person could build something equivalent.
Measured against that bar, most of the famous "open" models are not open source. They are open-weight, published under a custom license the company wrote itself, usually with conditions attached.
A smaller group of projects genuinely clears the bar — releases like the Allen Institute's OLMo family publish data, code and weights together specifically so the whole pipeline can be reproduced. Those are the exception, not the norm.
Why the wording matters to you: "open source" implies you can do what you like. "Open weight" only tells you the file is downloadable. The license tells you the rest, and the two are often very different.
The open-weight families you will meet
These are the names you will see on model hubs and in model pickers. Licensing changes from one release to the next, so treat this as orientation and always read the license on the specific model page.
| Family | Published by | Typical license style | Worth knowing |
|---|---|---|---|
| Llama | Meta | Custom community license | Commercial use allowed within limits; naming and attribution rules for derivatives |
| Gemma | Custom terms of use | Commercial use allowed; tied to a prohibited-use policy that follows the model downstream | |
| Qwen | Alibaba | Often Apache 2.0, some custom | Strong multilingual coverage; license varies by model |
| Mistral / Mixtral | Mistral AI | Mixed: Apache 2.0 and research-only | Some weights are fully permissive, others are non-commercial — check each one |
| DeepSeek | DeepSeek | Permissive on recent releases | Known for reasoning-focused models |
| Phi | Microsoft | Permissive | Small models aimed at modest hardware |
| Granite | IBM | Apache 2.0 | Enterprise-oriented and documentation-heavy |
| gpt-oss | OpenAI | Apache 2.0 | OpenAI's open-weight release line |
| OLMo | Allen Institute for AI | Permissive, with data and code | The closest thing to genuinely open-source AI |
Two practical notes I’d keep in mind. First, a license travels with the model: a community-quantized copy of Llama is still under Meta's license, no matter who uploaded it. Second, many of these downloads are gated — you click to accept the terms before the files unlock, which is itself a hint that there are terms worth reading.
The license clauses to check before you download
You do not need a lawyer to skim a model license. You need answers to six questions. For my own project, Xiar V0, the license column matters just as much as the benchmark column — a model I cannot ship is not a candidate, however well it scores.
- Is commercial use allowed at all? Plenty of impressive weights are released research-only, or under a non-commercial Creative Commons license. Fine for learning, not for a product.
- Are there conditions tied to your size? Some community licenses add thresholds — once a service passes a certain scale, the terms change and a separate agreement is required. Most people never touch this; a few need to know it exists.
- What happens if you fine-tune it? Derivative models often inherit the license and come with naming or attribution rules, so a tuned model may need to carry the original family name and a "built with" credit.
- Can you use the outputs freely? This is the one I see people miss. Some licenses restrict using a model's generated text to train other models.
- Is there an acceptable use policy attached? Custom licenses frequently point at a separate policy listing prohibited uses, and that policy binds you and anyone you pass the model on to.
- Can you redistribute the weights? If you plan to bundle a model inside a desktop app or a container image, you are redistributing. Check that this is permitted, and what notices you have to include.
If a license is short, says Apache 2.0 or MIT, and points nowhere else, your life is simple. If it is a custom document with links to other policies, read those links.
How to choose an open-weight model
Work through these in order. The list is deliberately short, because the first two steps eliminate most candidates.
1. Start with the license, not the leaderboard
Decide what you are building — a personal experiment, an internal tool, a product you will sell — and filter to models whose terms allow it. This saves you from falling in love with weights you cannot use.
2. Match the size to your hardware
Model size is quoted in billions of parameters: 3B, 8B, 70B. Quantization shrinks the file by storing those numbers at lower precision, which is how an 8B model fits comfortably on a mid-range graphics card. As a rough planning rule, a common 4-bit quantization needs somewhere around half a gigabyte of memory per billion parameters, plus headroom for the conversation itself.
A model that does not fit in your GPU memory spills into system RAM and crawls. I’d start smaller than you think you need and scale up. My guide to running an AI chatbot locally covers the hardware side in detail.
3. Pick the right variant
Most families ship several versions of the same release:
- Base — a raw completion model. Only useful if you intend to fine-tune it.
- Instruct / chat — tuned to follow instructions and hold a conversation. This is the one you want.
- Reasoning — trained to work through problems step by step. Slower, better at logic and code.
- Vision / multimodal — accepts images alongside text.
4. Check context length and languages
Context length is how much text the model can hold at once. Long documents and long conversations need a long context window, and memory use grows with it. If you need a language other than English, pick a family that advertises it rather than hoping.
5. Confirm the ecosystem supports it
A brilliant model with no convenient packaging is a weekend project. Before committing, check that it exists in a format your tools accept, and that the runner you use already lists it — my Ollama beginner's guide shows what that looks like in practice.
6. Test it on your own work
Benchmarks measure averages across tasks that are not yours. A good first test is to take five real prompts from the work you actually do, run them through two or three candidates at the same settings, and compare the answers side by side. That half-hour tells you more than any leaderboard.
Where open-weight models fit — and where they don't
Open weights give you privacy, no metered billing, the ability to work offline, and a model that cannot be deprecated out from under you. They usually cost you some capability at the very top end, plus the work of maintaining your own setup. That trade is one I’d take for anything I want to still work in five years — and I weighed both sides in local AI chatbot vs. ChatGPT.
Frequently Asked Questions
Are open-weight models free?
The download is almost always free, and nothing is metered while you run one. "Free" is not the same as "unrestricted", though — a free download can still carry a license that forbids commercial use. You also pay in hardware and electricity instead of per token.
Is Llama open source?
Not by the Open Source Initiative's definition. Llama is open-weight: the parameters are published and commercial use is permitted within limits, but it ships under Meta's own community license with conditions attached, rather than a standard open-source license.
Can I fine-tune an open-weight model and sell the result?
Often yes, if the license permits commercial use — but the derivative usually inherits that license, and some families require you to keep the original name in yours and credit the base model. Read the derivative-works section before you build a business on it.
Do open-weight models need a graphics card?
Not strictly. Smaller models run on a modern CPU with enough system RAM, just slowly. A GPU with 8 GB or more of memory is what makes a quantized mid-size model feel responsive rather than sluggish.
What is the difference between open weights and open data?
Open weights means the trained parameters are published. Open data means the training material is documented or released as well. Most open-weight releases publish the first and not the second, which is exactly why the open-source label does not fit them.
How do I check a model's license?
Open its page on the hub hosting it, look for the license field and the license file, then follow any linked acceptable-use or prohibited-use policy. If the model is a community re-upload, trace it back to the original publisher — the original terms still apply.
The takeaway
Open-weight models are downloadable, modifiable and yours to run; they are not automatically open source, and the license is where that difference shows up. Decide what you need to be allowed to do, filter to models that permit it, match the size to the hardware you own, then test the survivors on your own prompts.
Next step: pick one family from the table above, open its license page, and read it end to end. It takes ten minutes, and it will shape every decision you make after it.