What's the difference between an open-weight and an open-source AI model?

Open-weight means the model's trained parameters are downloadable, but the training data, code, or license freedoms may be withheld. Open-source (by the OSI definition) additionally requires data information and training/inference code under open terms. If the training data isn't disclosed, the model is open-weight, not open-source.

Is Llama open source?

No — Llama is open-weight under a community license. Meta doesn't release the training data, and the license adds an acceptable-use policy, a 700-million-monthly-active-user threshold, and restrictions on using outputs to train competing models. The OSI and others classify it as source-available or open-weight, not open source.

What is the OSI Open Source AI Definition?

Published in October 2024, it's the first stable definition of open-source AI. It grants four freedoms — use, study, modify, and share — and requires three components: data information, complete training and inference code, and the model parameters. Notably it requires detailed information about the training data rather than the raw dataset itself.

Does open-source AI require releasing the training data?

Not the raw dataset — the OSI definition requires sufficiently detailed data information so a skilled person could build a substantially equivalent system, and the actual data only when legally possible. This compromise is the most debated part of the definition, since true reproducibility arguably needs the data itself.

Can I use open-weight models commercially without paying?

Often, but it depends entirely on the license. Apache-2.0 models (many Qwen and Mistral releases) allow unrestricted commercial use. Llama's community license allows commercial use but with a user-count cap and other restrictions. Some variants are non-commercial only. Always check the specific model's license before deploying commercially.

Which AI models are truly open source?

Models that release weights plus the training data and the full training and inference code — OLMo from Ai2 is the leading example, publishing its corpus, code, logs, and checkpoints. Most popular "open" models (Llama, Qwen, Mistral, DeepSeek) are open-weight, withholding at least the training data.

Is an Apache-2.0 model the same as open source?

Apache-2.0 is a genuine open-source license, so Apache-2.0 weights are commercially unrestricted and permissive. But a model under Apache-2.0 weights still isn't fully "open source" by the OSI AI definition unless the training data information and code are also released — so it's best described as permissively licensed open-weight.

← Resources

By Ren Sugaya — Performance Engineer, GPU & Inference · 2026-04-12 · INSIGHT

Open-Weight vs Open-Source AI Models: What's the Difference?

Most models called "open source" are actually open-weight — the weights are downloadable but the training data and full license freedoms aren't. This explains the distinction, the OSI Open Source AI Definition, where popular models fall on the spectrum, and why the label matters for commercial use.

The core distinction

"Open-weight" means the trained model parameters are downloadable, but the training data, training code, or full license freedoms may be withheld. "Open-source AI," by the official definition, additionally requires data information and training and inference code, all under open terms. Without training-data disclosure, a release is open-weight, not open-source.

This matters because the terms are constantly conflated. A model you can download and run is not automatically "open source" in the way the software world means it — and the difference shows up exactly where it counts: commercial rights, reproducibility, and what you're actually allowed to do.

The OSI Open Source AI Definition

In October 2024 the Open Source Initiative published the first stable Open Source AI Definition. It grants four freedoms — to use the system for any purpose without permission, to study how it works, to modify it for any purpose, and to share it with or without modifications — and requires three components: data information (enough detail to build a substantially equivalent system), the complete training and inference code, and the model parameters.

Its most contested compromise is that it doesn't require releasing the raw training dataset itself — only sufficiently detailed information about it. Critics call that a weakening; defenders note that legal and privacy constraints often make releasing the raw data impossible. Either way, it's the bar against which "open source" claims should be measured.

Where popular models actually fall

Most well-known "open" models are open-weight, not open-source:

Llama is open-weight under a community license — Meta doesn't release the training data, and the license adds an acceptable-use policy, a 700-million-monthly-active-user clause, and restrictions on using outputs to train competing models. It fails the "use for any purpose without permission" freedom, so it's source-available, not open source. (See our [Llama guide](/articles/run-llama-locally-hardware-guide).)
Qwen and Mistral ship many models under Apache-2.0 — a true open-source license on the weights, so permissive commercial use, though without full data and code they're still "permissively licensed open-weight" rather than OSAID-open. (See our [Qwen3 guide](/articles/run-qwen3-locally-guide).)
OLMo (from Ai2) is the reference for truly open: weights plus the full training corpus, training code, logs, and checkpoints.
DeepSeek is open-weight — weights released, training data not disclosed.

The license spectrum

It helps to think of a spectrum from most to least open:

Fully open / OSAID-aligned: weights plus data plus code (OLMo).
Permissive OSI license on weights: Apache-2.0 (Qwen, Mistral) — commercially unrestricted weights, but not full data/code disclosure.
Restrictive "community" / source-available: Llama — usable commercially but with caps, acceptable-use rules, and competitor restrictions.
Non-commercial / research-only: some model variants restrict to non-commercial use.
API-only / closed: weights never released.

Knowing where a model sits tells you what you can legally do with it — which is the practical question, more than whether it earns the "open source" badge.

Why it matters, and "open-washing"

The label has real consequences. License terms — monthly-active-user caps, field-of-use bans, no-distillation clauses — directly determine commercial legality. Open-source (in the strict sense) additionally enables reproducing and auditing a model, which matters for trust and research. And regulation raises the stakes: the EU AI Act gives open-source systems certain carve-outs, so calling an open-weight model "open source" isn't just sloppy — it can be consequential.

That's the heart of the "open-washing" critique: marketing a model as open source when it's open-weight at best. Studies of generative-AI systems have found many "open source" labels are open-weight in reality, and the broader trend confirms it — the share of model downloads that disclose training data has fallen sharply, meaning the ecosystem is going open-weight, not open-source.

How osFoundry treats it

osFoundry runs open-weight models regardless of where they sit on the spectrum — permissive Apache-2.0 (Qwen, Mistral), restrictive community-licensed (Llama), and fully open (OLMo) all run the same way through BYOK or self-hosting. Because the open-versus-open-weight question is fundamentally about licensing, osFoundry surfaces each model's actual license terms — the user-cap clauses, no-distillation rules, and non-commercial restrictions — so teams can confirm commercial legality before deploying. With BYOK and self-host, you keep the weights, data, and inference under your own control rather than behind a vendor API, which is the practical payoff of "open" whether or not a given model clears the strict open-source bar.

Frequently asked questions

What's the difference between an open-weight and an open-source AI model?: Open-weight means the model's trained parameters are downloadable, but the training data, code, or license freedoms may be withheld. Open-source (by the OSI definition) additionally requires data information and training/inference code under open terms. If the training data isn't disclosed, the model is open-weight, not open-source.
Is Llama open source?: No — Llama is open-weight under a community license. Meta doesn't release the training data, and the license adds an acceptable-use policy, a 700-million-monthly-active-user threshold, and restrictions on using outputs to train competing models. The OSI and others classify it as source-available or open-weight, not open source.
What is the OSI Open Source AI Definition?: Published in October 2024, it's the first stable definition of open-source AI. It grants four freedoms — use, study, modify, and share — and requires three components: data information, complete training and inference code, and the model parameters. Notably it requires detailed information about the training data rather than the raw dataset itself.
Does open-source AI require releasing the training data?: Not the raw dataset — the OSI definition requires sufficiently detailed data information so a skilled person could build a substantially equivalent system, and the actual data only when legally possible. This compromise is the most debated part of the definition, since true reproducibility arguably needs the data itself.
Can I use open-weight models commercially without paying?: Often, but it depends entirely on the license. Apache-2.0 models (many Qwen and Mistral releases) allow unrestricted commercial use. Llama's community license allows commercial use but with a user-count cap and other restrictions. Some variants are non-commercial only. Always check the specific model's license before deploying commercially.
Which AI models are truly open source?: Models that release weights plus the training data and the full training and inference code — OLMo from Ai2 is the leading example, publishing its corpus, code, logs, and checkpoints. Most popular "open" models (Llama, Qwen, Mistral, DeepSeek) are open-weight, withholding at least the training data.
Is an Apache-2.0 model the same as open source?: Apache-2.0 is a genuine open-source license, so Apache-2.0 weights are commercially unrestricted and permissive. But a model under Apache-2.0 weights still isn't fully "open source" by the OSI AI definition unless the training data information and code are also released — so it's best described as permissively licensed open-weight.