What Does “Open Source LLM” Actually Mean? Weights, Code and Commercial Licences

An LLM is not automatically open source because its weights can be downloaded. Weights, inference code, training information, datasets and service access have separate permissions. For practical decisions, identify what is available, what you are allowed to do with it and which obligations remain when you modify or distribute the system.

Use precise labels

Open weights generally means model parameters are accessible under specified terms. It tells you something about availability, but not necessarily unrestricted commercial use or access to the information needed to study and change the whole system.

Open-source inference software is code distributed under an open-source licence. You can run such an engine with a model whose weights have more restrictive conditions. The engine's licence does not override the model's terms, just as a media player's licence does not determine ownership of every file it opens.

A free hosted service is different again. You may receive limited access without possessing weights or permission to redistribute anything. Its price, usage limits and data rules can change independently of an underlying model's availability.

The Open Source Initiative's AI Definition sets out freedoms to use, study, modify and share an AI system, together with access requirements for making modifications. It is a useful framework, but a project's marketing label should still be checked against its actual released components and terms.

Read the licence as an operational requirement

Start with the exact model card and licence file, not a search snippet. Save the checkpoint revision and the text or version of the terms you reviewed. A publisher may use different terms for separate families or sizes.

Ask whether commercial use, redistribution, derivative models and hosted access are permitted. Check attribution, notices, acceptable-use conditions and any special restrictions relevant to your planned product. If a clause is unclear for a significant business decision, seek a qualified review rather than guessing.

Derived models need additional attention. An adaptation may involve base weights, a fine-tuning dataset, added code and a separate conversion. Trace each relevant component. The newest uploader cannot grant rights they did not receive from upstream authors.

For example, the DeepSeek-R1 repository explicitly identifies differing original licences behind its Qwen-derived and Llama-derived distilled checkpoints. That illustrates why “all models with this brand have the same licence” is an unsafe shortcut.

Keep software and model notices together in the deployment record. A permissive model licence does not automatically resolve rights to training data, trademarks, personal information or generated material. Those are related but distinct questions.

Build a reusable selection record

For each candidate, record the original publisher, model identifier, revision, source location, weight licence, runtime licence and intended use. Add the person responsible for reviewing updates. This can be a short document rather than a complicated compliance system.

Then write down how the model will be used. Internal research, a public chatbot and redistributing modified weights create different questions. A vague statement such as “we may use AI later” is insufficient to evaluate relevant conditions.

Check whether a download requires accepting additional terms. Do not assume a model browser's one-word licence badge captures every linked document. Read the original files and retain the source links so another reviewer can reproduce the decision.

If you publish a comparison, use separate columns for weight availability, software licence and commercial conditions. Avoid a single yes-or-no “open source” badge that hides meaningful distinctions. Where evidence is incomplete, say precisely what has not been verified.

Your next step is to select one candidate and trace its rights from publisher to deployed artefact. This habit is more useful than memorising a list of supposedly unrestricted models, because releases, distribution routes and intended uses all change.

Quick questions, clear answers

Does Apache 2.0 mean I can ignore notices?

No. Permissive licences still have conditions, including relevant notice and licence obligations. Read the actual licence and any accompanying notices. Also confirm that the model you downloaded is genuinely covered by those terms rather than assuming every artefact from the same organisation is identical.

Can I call a free API open source?

Free access describes pricing, while open source concerns rights and availability of the relevant system components. A service can provide free usage without releasing code or weights. Describe it as a free service unless you have evidence supporting a more specific claim.

Why distinguish open source from open weights publicly?

Readers need to understand whether they can inspect, modify, host and redistribute a system. Precise labels prevent a developer from making a deployment decision based on permissions that were never granted. They also make comparisons between projects clearer without dismissing useful open-weight releases.

About this guide

Aniket Thakur prepared this guide from secondary research checked on 22 October 2026. No first-hand software testing or performance benchmarking was conducted. Examples are instructional and require validation in your environment.