Open Weight vs Open Source AI

Who this is for

Beginners who keep seeing "open source" on model cards and want to know what it commits the publisher to, plus founders, developers and teams recording licences before they ship.

Recommended stack

• The OSI Open Source AI Definition (OSAID 1.0) as the yardstick for the phrase "open source AI"
• The LICENSE file and any usage-policy file in the repository the weights are served from, not the announcement post
• A four-field licence note per model: source, licence string, usage policy, commercial terms
• This site's model pages, which show the licence string from the catalog record

Start with the word "open": it hides three different claims

When someone says a model is open, they may mean one of three things, and the three are not interchangeable. Open weights means the trained parameters (the numbers that make the model work) can be downloaded and run on your own machine, under whatever licence the publisher chose. Openly licensed weights means those parameters come under a licence the Open Source Initiative (OSI) has approved for software, such as Apache 2.0 or MIT, so you can use them for any purpose, including commercially, without asking. Open Source AI is a claim about the whole system, not just the weights, and it has a written definition. Most models described as open today are the first kind, a good number are the second, and almost none are the third. Once you can tell which claim you are looking at, most of the confusion in this area goes away.

What Open Source AI means: the OSI definition

The Open Source Initiative published the Open Source AI Definition (OSAID 1.0) to say precisely what the phrase should mean for AI systems, the same way its Open Source Definition has done for software for decades. An AI system counts as Open Source AI only if it grants four freedoms (use it for any purpose without asking permission, study how it works, modify it, and share it with or without changes) and makes available, under OSI-approved terms, what OSAID calls the preferred form to make modifications. That preferred form has three parts. Data information: sufficiently detailed information about the data used to train the system (its provenance, and how it was labelled and filtered) that a skilled person could build a substantially equivalent system, though the data itself does not have to be shared. Code: the complete source code used to train and run the system, covering data processing, training, validation, inference and architecture. Parameters: the weights themselves, including intermediate checkpoints and optimizer states. OSI does not publish a list of compliant models; the definition is a test you apply, not a badge someone hands out.

Why open weights alone do not clear that bar

A weights-only release gives you the third of those three parts. That is genuinely useful: you can run the model offline, fine-tune it, quantize it and ship it, and this whole site is built around doing exactly that. But without the training code and the data information, you cannot study how the model came to be or rebuild a substantially equivalent one, which is what the study and modify freedoms in OSAID are meant to protect. So the honest label for a downloadable model under a permissive licence is openly licensed weights, or simply open weights, rather than open source AI. This is not a criticism of the publishers; releasing weights at all is a real contribution to the people who use this site. It is just a narrower claim than the phrase open source AI implies, and the difference matters when you are deciding what you can rely on.

The licence on the weights: four plain-language classes

The licence attached to the weights is a separate question from whether the system meets OSAID, and it is the one that most affects what you may do day to day. This site sorts model licences into four plain-language classes. OSI-approved licence: Apache 2.0, MIT and similar. Commercial use is allowed and the licence itself has no field-of-use limits, though a publisher may still ship a separate usage policy next to the weights (more on that below). Custom or community licence with restrictions: the Llama 3, 3.1, 3.3 and 4 Community Licenses, the Gemma Terms of Use, and the Qwen License on older Qwen 2.5 checkpoints. (The GLM License was listed here until rev 198; every GLM record in both catalogs is MIT, and the one that said otherwise was wrong.) These are usable, often commercially, but they carry acceptable-use policies and sometimes scale thresholds, so you have to read them. Non-commercial: for example the MiniMax M2 non-commercial licence and various research-only licences; fine for learning and evaluation, not for a product. Proprietary or hosted only: no weights to download; you use the model through someone else's API on their terms.

Where the plainest licences in this catalog come from

It is worth noticing that some of the most permissively licensed weights in this catalog come from labs in China: Alibaba's Qwen3 family ships under Apache 2.0 and DeepSeek's R1 distills under MIT, both plain OSI-approved licences with no community terms attached. That cuts against the assumption that a familiar brand means a cleaner licence, or that a less familiar one means a murkier one. Read the licence string on the record, not the logo on the announcement.

Software versus weights: the Grok example

The same publisher can be open in one layer and closed in another, and Grok is a clean illustration. Grok Build, xAI's terminal coding agent, is software: its harness is released under Apache 2.0, so calling that harness open source is accurate in the ordinary software sense. The model it talks to by default is hosted and metered, so nothing about that model is open at all. Meanwhile xAI's older grok-1 weights are under Apache 2.0, and the later grok-2 weights sit under a more restrictive community licence. Four things called Grok, three different answers. When you evaluate a tool, ask separately about the software you install and the model it runs, because the licence on one tells you nothing about the other.

Is Llama open source? What OSI has said

Meta describes Llama as open source, and OSI has twice explained in writing why it disagrees: once for Llama 2 in 2023, and again in 2024 in a post titled, plainly, that the licence is still not open source. The Llama Community License restricts commercial use for very large services and attaches an acceptable-use policy. Those are limits on who may use the model and for what, and the Open Source Definition forbids discriminating against fields of endeavour or against persons and groups. Llama is therefore open weights under a custom community licence, and it is a perfectly reasonable model to run locally within those terms. It just is not open source in the sense OSI defines, and this site labels it accordingly.

Is Muse Glimmer open source? Apache 2.0 weights, a usage policy, no training recipe

Muse Glimmer 30B, released by Meta on 10 August 2026, is a useful test of everything above, because it is the only Meta model in this catalog whose weights carry a plain OSI-approved licence. The weights are Apache 2.0. The repository also carries a separate USAGE_POLICY.md file, and the model card describes the training data only as publicly available data, data provided by third parties and information from Meta's products and services, with no training code or data recipe published. OSI's executive director wrote on 11 August 2026 that the licence is Apache 2.0, which is an open-source licence, and asked what the usage policy does; he made no claim that the model meets the Open Source AI Definition. The accurate description is therefore: open weights under an OSI-approved licence, and it does not meet OSAID, because the data information is minimal and the training code is unavailable. Both halves of that sentence are true at once, and neither cancels the other.

Open source software is a different question from open source AI

Everything above is about models. For the software around them (runtimes, chat interfaces, agent frameworks, vector databases), the ordinary Open Source Definition applies and the picture is simpler: either the licence is OSI-approved or it is not. Two phrases to know. Source-available means you can read the code but the licence restricts what you may do with it; the Elastic License and the fair-code family are common examples, and this site labels such tools source-available rather than open source. Open core means a genuinely open-source core with proprietary paid layers on top. A tool can be fully open source while the model it runs is proprietary, and the other way round; the Grok example above is exactly that pairing.

Where the models in this catalog stand

The next five sections were hand-checked against the catalog behind the Compatibility Checker on 16 August 2026, grouping the 33 checker-visible models by the exact licence string on their record. They state the licence class of the weights only. As far as this site has been able to check, none of these publishers has released the training code together with the data information that OSAID asks for, so no model on this list is claimed here to meet the Open Source AI Definition; that includes the ones under Apache 2.0 and MIT. A structured per-model badge that reads this from the catalog automatically is planned; until it ships, treat this list as the reference and each model page as the place to confirm the string.

In this catalog: Apache 2.0 (23 of 35)

Apache 2.0, an OSI-approved licence: Muse Glimmer 30B (Meta); GPT-OSS 20B (OpenAI); Mistral Small 3.1 (Mistral AI); Gemma 4 12B (QAT), Gemma 4 26B (A4B) and Gemma 4 31B (Google); Qwen3 4B, Qwen3 8B, Qwen3 14B, Qwen3 32B, Qwen3 30B (A3B), Qwen3-Coder 30B (A3B), Qwen3.5 9B, Qwen3.6 27B, Qwen3.6 35B (A3B), Qwen3.8 27B, Qwen2.5 7B, Qwen2.5 14B, Qwen2.5 32B, Qwen2.5-Coder 7B, Qwen2.5-Coder 14B and Qwen2.5-Coder 32B (Alibaba); North Mini Code 1.0 (Cohere Labs). Commercial use is permitted by the licence itself. Muse Glimmer 30B additionally ships a usage-policy file beside the weights, so read both documents.

In this catalog: MIT (4 of 35)

MIT, an OSI-approved licence: Phi-4 and Phi-3 Mini 128K (Microsoft); DeepSeek-R1 8B (0528 Qwen3) and DeepSeek-R1 Distill 14B (DeepSeek). MIT is shorter than Apache 2.0 and, unlike Apache 2.0, includes no express patent grant; for running a model on your own hardware the practical permissions are the same.

In this catalog: Gemma Terms of Use (3 of 35)

Gemma Terms of Use, a custom licence: Gemma 3 4B, Gemma 3 12B and Gemma 3 27B (Google). Usable, including commercially, but with an acceptable-use policy attached, so read it. The newer Gemma 4 records in this catalog carry Apache 2.0 instead, which is why two generations of the same family sit in different groups here, and a good reminder that the licence belongs to the checkpoint, not the brand.

In this catalog: Llama Community Licenses (4 of 35)

Each Llama generation carries its own licence string, so they are listed separately. Llama 3 Community License: Llama 3 70B. Llama 3.1 Community License: Llama 3.1 8B Instruct. Llama 3.3 Community License: Llama 3.3 70B. Llama 4 Community License: Llama 4 Scout (all Meta). All four are custom community licences with an acceptable-use policy and a scale threshold on commercial use, which is why OSI does not consider them open source. They are open weights, and the terms differ between generations, so read the version that matches your checkpoint.

In this catalog: OpenMDW-1.1 (1 of 35)

OpenMDW-1.1: Laguna XS 2.1 (Poolside). OpenMDW (Open Model, Data and Weights) is a permissive licence written specifically for models, distinct from the OSI-approved software licences above, so it does not fit neatly into the four classes. Its record on this site says the same thing: check the terms before commercial deployment.

What to record for every model you use

Four fields cover almost every question you will get later. Source: where the weights came from, down to the exact repository and revision, since the same name can point at different checkpoints. Licence: the exact string, copied from the LICENSE file rather than the announcement. Usage policy: whether a separate acceptable-use or usage-policy file ships alongside the weights, and what it prohibits. Commercial terms: whether commercial use is allowed, and whether any threshold, attribution line or notice requirement applies. Keep these in your model registry, or a spreadsheet if that is what you have; the point is that the answer exists in one place before a customer, an investor or a lawyer asks.

What "check the model card" really means

Advice to check the model card appears all over this site, so here is what it means in practice. Open the repository the weights are actually served from, not a mirror. Read the LICENSE (or LICENSE.txt) file itself: the summary line on the model card can be out of date, and a repository can carry a licence and a separate usage policy in two different files. Look for a training-data section and notice whether it names sources or only describes them in general terms; that difference is exactly what separates open weights from the data information OSAID asks for. Confirm the revision you are downloading, because licences have changed between checkpoints of the same family. Five minutes here saves a lot of guessing later.

When to ask a lawyer

You do not need counsel to learn, prototype or evaluate; that is what permissive and community licences are designed to allow. Ask for a professional read before you ship a model to paying customers, when a licence contains a scale or revenue threshold and your product might grow into it, when a usage policy is broad enough that you cannot tell whether your use case is inside it, when you plan to redistribute or sell fine-tuned weights, or when a licence is non-commercial or research-only and you have any commercial intent at all. This guide explains the vocabulary; it is not legal advice.

Practical recommendations

• Say open weights, or openly licensed weights, unless the publisher has released training code and data information under OSI-approved terms.
• Record source, licence string, usage policy and commercial terms for every model before it reaches a product.
• Read the LICENSE file and any usage-policy file in the repository, not the announcement post.
• Prefer OSI-approved licences (Apache 2.0, MIT) when commercial use matters, and still read any usage policy that ships beside the weights.
• Ask counsel before shipping under a non-commercial, research-only or threshold-bearing licence.

Tradeoffs

The most capable model for your task may carry the least convenient licence, and the model with the cleanest licence may still publish nothing about how it was trained. You usually have to decide which kind of openness matters most for the job in front of you, and record what you chose.

Related links

FAQ

Sources