CyberFreezeDevPermissive
Apache 2.0 and MIT licences allow commercial use with few conditions.
Read more about Permissive
Apache 2.0 and MIT licences (used for some Mistral, Qwen, DeepSeek and gpt-oss releases) allow commercial use with few conditions.
- Home
- AI glossary
- Open-weight model
What is an open-weight model?
Models whose learned weights you can download and run on your own machines.
Definition
The developer publishes the weights so anyone can run them, unlike API-only models. Open-weight is not always open source: training data is usually not shared.
Read more about Definition
Every neural network is a set of weights: billions of numbers learned during training. An open-weight model is one where the developer publishes those weights, usually on Hugging Face, so anyone can load them into their own software. Examples include Meta's Llama, Alibaba's Qwen, Mistral's models, Google's Gemma, DeepSeek's models and OpenAI's gpt-oss.
The opposite is a closed or API-only model, which you can use only by sending requests to the provider's servers.
"Open-weight" is not always the same as "open source". Open source in the strict sense also means the training code and enough information about the data to rebuild the model, as set out in the Open Source Initiative's Open Source AI Definition. Most open-weight releases share the weights but not the training data.
Licences differ, read them
Custom community licences
Llama allows commercial use with conditions; Gemma has its own terms and prohibited uses.
Read more about Custom community licences
Llama's licence allows commercial use but adds conditions, such as attribution and a separate licence for very large platforms. Gemma has its own terms of use with a prohibited use policy.
Research only
Some releases forbid commercial use entirely. Check before building on one.
Read more about Research only
Some releases forbid commercial use entirely. Check before building a product on one.
Where it matters in business
Data control
Run it on your own servers so client data never reaches an outside AI provider.
Read more about Data control
Run the model on your own servers so client data never reaches a third-party AI provider.
Customisation
Fine-tune or train a LoRA adapter, which closed models allow only in limited ways.
No forced upgrades
The model you tested stays the model you run; nobody can change or retire it.
Read more about No forced upgrades
The model you tested is the model you run. A provider cannot change or retire it under you.
A practical example
A company tests three open models on fifty real questions, checks licences, picks the smallest that passes, and reruns the test before any switch.
Read more about A practical example
A company wants an assistant that drafts replies from its policy manuals. It tests three open-weight models of different sizes against fifty real questions with known answers, checks their licences for commercial use, picks the smallest one that passes, quantizes it to fit a single GPU, and serves it inside the office network. When a better model is released, it reruns the same fifty questions before switching.
Common mistakes
Assuming open means free of obligations
Licences can restrict uses, require attribution or limit very large deployments.
Read more about Assuming open means free of obligations
Licences can restrict use cases, require attribution or limit very large deployments.
Downloading from unknown uploaders
Use the official repository or a well-known converter; model files can be tampered with.
Read more about Downloading from unknown uploaders
Use the developer's official repository or a well-known converter. Model files can be tampered with.
Skipping evaluation
Benchmark scores are not your tasks. Evaluate on your own data.
Read more about Skipping evaluation
Benchmark scores are not your tasks. Run a model evaluation on your own data.
Related terms and pages
Questions
Are open-weight models as good as ChatGPT or Claude?
The largest open-weight models are competitive on many tasks, but the models you can run on modest hardware are smaller and weaker at complex reasoning. For narrow, well defined business tasks a smaller open model is often enough. Test on your own work.
Can I use an open-weight model commercially?
Many allow it, but licences vary. Apache 2.0 and MIT models are the simplest. Read the licence of the exact model and version before you build on it.
Is my data used to train the model if I run it myself?
No. Running a downloaded model on your own hardware sends nothing back to the developer unless your software is set up to do so.
Last updated