Open weights describe access to trained model parameters. Open source AI concerns the freedoms and materials needed to use, study, modify and share an AI system. A weight download alone does not answer the licensing or reproducibility questions.
Three labels that answer different questions
| Label | What it usually tells you | What you still need to check |
|---|---|---|
| API access | You can send requests to a hosted model. | Service terms, retention, availability and whether weights are available separately. |
| Open or downloadable weights | The trained parameters are available to download under stated terms. | Commercial use, redistribution, modification, use restrictions and available code or data information. |
| Open source AI | A claim about freedoms and the materials supporting them. | Which definition is being used, and whether the actual release satisfies it. |
“Open-weight” is useful descriptive shorthand, but it does not specify one standard license. Two models with downloadable files may permit different uses. Conversely, a hosted service can also offer a separately licensed downloadable model. Evaluate the specific release, not just the organization’s reputation or product family.
What does the OSI definition require?
The Open Source Initiative’s Open Source AI Definition 1.0 describes freedoms to use the system for any purpose, study how it works, modify it and share it. Its preferred form for modification includes parameters, relevant code and sufficiently detailed information about the training data.
This is more specific than saying that a model has “open weights”. The definition does not simply require every original training item to be downloadable: it specifies data information and distinguishes different kinds of training data. Read the definition itself when assessing a claim. Whether a particular release meets it needs a release-specific review.
A release checklist before you build
- Record the exact model and revision. Save the repository, checkpoint identifier and license version. A family name can cover releases with different terms.
- Read the license and any referenced policy. Check use, modification, distribution, attribution and conditions on derivative works. Do not infer commercial permission from the download button.
- Inspect the model card. Look for intended use, limitations, training information, evaluation conditions and supported languages. Missing information is an uncertainty to record.
- Check code and tokenizer requirements. Determine what must run locally and whether custom code is required. Weight availability does not guarantee compatibility with your preferred runtime.
- Check downstream artifacts. Quantized files, adapters and merged models can introduce additional provenance or compatibility questions. Confirm the base model and conversion process.
- Keep the evidence with your deployment. Retain license copies, model-card links and revision identifiers so a later update does not silently change your assumptions.
Downloadable does not mean easy to run
Model size is only one part of deployment. Memory also holds the KV cache, runtime buffers and other allocations. Long contexts and concurrent requests can materially increase requirements. Quantization reduces weight storage but may change numerical behavior, runtime support and speed.
Use the LLM memory calculator as an initial estimate, then test the exact model, format and workload. It is not a throughput benchmark or a guarantee that a checkpoint fits a particular GPU.
Local operation gives you control over where inference runs, but it does not automatically make the entire application private. Downloads, update checks, telemetry and connected tools can still communicate externally. Inspect the complete deployment and its settings.
How to compare releases fairly
Start with your task and constraints: required languages, accuracy, latency, memory budget and permitted use. Run the same evaluation set with comparable prompts and inference budgets. Keep licensing suitability separate from measured capability; a highly capable model may still be unsuitable for your distribution plan.
A model can offer useful access without meeting an open source definition. Precise language helps readers understand that tradeoff. Our open-model coverage follows releases and local-inference developments; consult the source release for its actual terms.
Sources and scope
- OSI: Open Source AI Definition 1.0 — the definition used in this comparison.
- Hugging Face: Model Cards — documenting model use, limitations, training and evaluation.
- Transformers quantization documentation — practical formats and supported integrations.
This guide compares concepts and supplies a reading checklist. It does not classify every released model or resolve a specific license interpretation. Report an error or suggest an update.