[{"data":1,"prerenderedAt":110},["ShallowReactive",2],{"learn-open-weight-vs-open-source":3},{"collection":4,"slug":5,"path":6,"title":7,"seoTitle":8,"description":9,"summary":10,"primaryKeyword":11,"cluster":12,"published":13,"updated":13,"reviewed":13,"readingMinutes":14,"html":15,"toc":16,"wordCount":39,"sources":40,"faq":74,"related":87,"tested":104,"internalLinks":105},"learn","open-weight-vs-open-source","\u002Flearn\u002Fopen-weight-vs-open-source\u002F","Open-weight vs open-source AI","Open-Weight vs Open-Source AI: What the License Allows","Open-weight means you can download the weights. Open source AI means meeting the OSI definition. What MIT, Apache 2.0, Llama and Gemma terms actually allow.","Open-weight means the trained parameters can be downloaded. Open source AI is a stricter claim: under the OSI Open Source AI Definition, the terms must let anyone use, study, modify and share the system, and the release must include data information, training code and the parameters. A model under MIT or Apache 2.0 can still fall short, and community licenses such as Llama's add use restrictions the definition does not allow.","open weight vs open source ai","licensing","2026-11-14",7,"\u003Ch2 id=\"two-separate-questions\">Two separate questions\u003C\u002Fh2>\n\u003Cp>“Open-weight” and “open source” answer different questions. The first is practical: can you download the trained parameters and run them yourself? The second is about rights and materials: do the terms let anyone use, study, modify and share the model, and did the developer publish what someone would need to rebuild it?\u003C\u002Fp>\n\u003Cp>Almost every model on this site is open-weight. The files are on Hugging Face and run on your own hardware. That tells you nothing about the license. GLM-5.3-Flash ships under MIT, Gemma 4 under Apache 2.0, Llama 3.3 under Meta’s own community license and Gemma 3 under Google’s Gemma Terms of Use. These licenses allow different things, and none of them, on its own, makes a model open source in the OSI sense.\u003C\u002Fp>\n\u003Ch2 id=\"what-the-osi-definition-requires\">What the OSI definition requires\u003C\u002Fh2>\n\u003Cp>The Open Source Initiative published version 1.0 of the \u003Ca href=\"https:\u002F\u002Fopensource.org\u002Fai\u002Fopen-source-ai-definition\" rel=\"noopener\" class=\"ext\">Open Source AI Definition\u003C\u002Fa> (OSAID). It starts from four freedoms. An Open Source AI must be available under terms that let anyone:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Use\u003C\u002Fstrong> the system “for any purpose and without having to ask for permission”;\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Study\u003C\u002Fstrong> how it works and inspect its components;\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Modify\u003C\u002Fstrong> it for any purpose, “including to change its output”;\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Share\u003C\u002Fstrong> it with others, with or without modifications, for any purpose.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>The definition then says what you need in hand to exercise those freedoms, which it calls the “preferred form to make modifications”. For a machine-learning system, that form must include all three of these elements.\u003C\u002Fp>\n\u003Cdiv class=\"table-wrap\" role=\"region\" aria-label=\"Scrollable table\" tabindex=\"0\">\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Element\u003C\u002Fth>\n\u003Cth>What the OSAID asks for\u003C\u002Fth>\n\u003Cth>Terms it must be under\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\n\u003Ctr>\n\u003Ctd>Data information\u003C\u002Ftd>\n\u003Ctd>“Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system”: provenance, selection, labeling and filtering, plus a listing of public and third-party training data and where to get it\u003C\u002Ftd>\n\u003Ctd>OSI-approved terms\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Code\u003C\u002Ftd>\n\u003Ctd>“The complete source code used to train and run the system”: data processing and filtering, training settings, validation, tokenizers, inference code and model architecture\u003C\u002Ftd>\n\u003Ctd>OSI-approved licenses\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Parameters\u003C\u002Ftd>\n\u003Ctd>“The model parameters, such as weights or other configuration settings”, which may include intermediate checkpoints and the final optimizer state\u003C\u002Ftd>\n\u003Ctd>OSI-approved terms\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\u003Cp>Three details are easy to miss.\u003C\u002Fp>\n\u003Cp>First, the data itself does not have to be published. The definition allows “unshareable data” as long as it is described. OSI’s \u003Ca href=\"https:\u002F\u002Fopensource.org\u002Fai\u002Ffaq\" rel=\"noopener\" class=\"ext\">FAQ\u003C\u002Fa> gives medical data and personally identifiable information as examples of data that cannot always be shared.\u003C\u002Fp>\n\u003Cp>Second, weights alone cannot qualify. The OSAID states that “Open Source models” and “Open Source weights” “must include the data information and code used to derive those parameters.”\u003C\u002Fp>\n\u003Cp>Third, copyleft is allowed. The terms may require that modified versions be released under the same terms as the original.\u003C\u002Fp>\n\u003Cp>OSI also reported which models passed its validation phase: Pythia, OLMo, Amber, CrystalCoder and T5. It named Llama 2, Grok, Phi-2 and Mixtral among the models analyzed that did not pass, and it adds that these results “are not certifications of any kind”. None of the base models covered on this site appears on either list, and we have not assessed any of them against the definition.\u003C\u002Fp>\n\u003Ch2 id=\"where-open-weight-releases-stop\">Where open-weight releases stop\u003C\u002Fh2>\n\u003Cp>OSI’s explainer on \u003Ca href=\"https:\u002F\u002Fopensource.org\u002Fai\u002Fopen-weights\" rel=\"noopener\" class=\"ext\">open weights\u003C\u002Fa> describes the gap: a weights release does not include the training code, the training dataset or, when the data cannot legally be shared, “comprehensive data transparency”. You can fine-tune, quantize or edit the weights. You cannot reproduce how they were made.\u003C\u002Fp>\n\u003Cp>This matters for refusal-reduced models. An \u003Ca href=\"\u002Flearn\u002Fwhat-is-abliteration\u002F\">abliterated\u003C\u002Fa> or fine-tuned variant changes weights whose training data nobody outside the original developer can inspect. You can measure how the variant behaves. You cannot trace why the base model behaved as it did.\u003C\u002Fp>\n\u003Ch2 id=\"what-each-license-lets-you-do-with-the-weights\">What each license lets you do with the weights\u003C\u002Fh2>\n\u003Cp>The license is a separate layer from the OSAID’s materials test. The table summarizes what the license texts say for the four licenses that cover most models on this site. It paraphrases; read the full text before relying on any row.\u003C\u002Fp>\n\u003Cdiv class=\"table-wrap\" role=\"region\" aria-label=\"Scrollable table\" tabindex=\"0\">\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Question\u003C\u002Fth>\n\u003Cth>MIT\u003C\u002Fth>\n\u003Cth>Apache 2.0\u003C\u002Fth>\n\u003Cth>Llama 3.3 Community License\u003C\u002Fth>\n\u003Cth>Gemma Terms of Use\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\n\u003Ctr>\n\u003Ctd>Used by\u003C\u002Ftd>\n\u003Ctd>GLM-5.3-Flash, GLM-4.7-Flash\u003C\u002Ftd>\n\u003Ctd>Gemma 4, Qwen3.8, gpt-oss-20b\u003C\u002Ftd>\n\u003Ctd>Llama 3.3\u003C\u002Ftd>\n\u003Ctd>Gemma 3 (also Gemma 1, 1.1, 2 and 3n)\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Commercial use\u003C\u002Ftd>\n\u003Ctd>Yes; the grant includes the right to “sell copies”\u003C\u002Ftd>\n\u003Ctd>Yes; no field-of-use limit in the text\u003C\u002Ftd>\n\u003Ctd>Yes, unless your products had more than 700 million monthly active users on the release date, in which case you must request a license from Meta\u003C\u002Ftd>\n\u003Ctd>Allowed within the terms; no user threshold in the text\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Restricted purposes\u003C\u002Ftd>\n\u003Ctd>None in the text\u003C\u002Ftd>\n\u003Ctd>None in the text\u003C\u002Ftd>\n\u003Ctd>Acceptable Use Policy, incorporated by reference\u003C\u002Ftd>\n\u003Ctd>Prohibited Use Policy, incorporated by reference\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Redistributing modified weights\u003C\u002Ftd>\n\u003Ctd>Keep the copyright and permission notice\u003C\u002Ftd>\n\u003Ctd>Include the license, mark changed files, keep notices and any NOTICE file\u003C\u002Ftd>\n\u003Ctd>Include the agreement, display “Built with Llama”, keep the Notice text, start the model’s name with “Llama”\u003C\u002Ftd>\n\u003Ctd>Pass on the use restrictions as enforceable terms, include the agreement, mark modified files, add a Notice file\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>Serving it to others as a service\u003C\u002Ftd>\n\u003Ctd>Not addressed\u003C\u002Ftd>\n\u003Ctd>Not addressed\u003C\u002Ftd>\n\u003Ctd>Duties apply if you “distribute or make available” it, including in a product or service\u003C\u002Ftd>\n\u003Ctd>A hosted service counts as “Distribution”\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>On OSI’s approved list\u003C\u002Ftd>\n\u003Ctd>Yes\u003C\u002Ftd>\n\u003Ctd>Yes\u003C\u002Ftd>\n\u003Ctd>No\u003C\u002Ftd>\n\u003Ctd>No\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\u003Cp>\u003Ca href=\"https:\u002F\u002Fopensource.org\u002Flicense\u002Fmit\" rel=\"noopener\" class=\"ext\">MIT\u003C\u002Fa> and \u003Ca href=\"https:\u002F\u002Fopensource.org\u002Flicense\u002Fapache-2-0\" rel=\"noopener\" class=\"ext\">Apache 2.0\u003C\u002Fa> both appear on OSI’s \u003Ca href=\"https:\u002F\u002Fopensource.org\u002Flicenses\" rel=\"noopener\" class=\"ext\">list of approved licenses\u003C\u002Fa>. Both were written for software, and OSI’s FAQ notes that the definition “does not take a stance on the legal nature of Parameters”.\u003C\u002Fp>\n\u003Cp>For Llama, OSI wrote in 2023 that \u003Ca href=\"https:\u002F\u002Fopensource.org\u002Fblog\u002Fmetas-llama-2-license-is-not-open-source\" rel=\"noopener\" class=\"ext\">Meta’s Llama 2 license “is not Open Source”\u003C\u002Fa>. It pointed to two things: the commercial limit for some users and the use restrictions in the Acceptable Use Policy. The \u003Ca href=\"https:\u002F\u002Fwww.llama.com\u002Fllama3_3\u002Flicense\u002F\" rel=\"noopener\" class=\"ext\">Llama 3.3 license\u003C\u002Fa> keeps both, in Section 2 and Section 1.b.iv.\u003C\u002Fp>\n\u003Cp>The \u003Ca href=\"https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms\" rel=\"noopener\" class=\"ext\">Gemma Terms of Use\u003C\u002Fa> have no user threshold. They do incorporate a Prohibited Use Policy, and Google “reserves the right to restrict (remotely or otherwise) usage” that it reasonably believes breaches the terms.\u003C\u002Fp>\n\u003Ch2 id=\"a-decision-list-for-your-use-case\">A decision list for your use case\u003C\u002Fh2>\n\u003Cp>Work through these questions in order, for the base model of whatever you plan to run. A refusal-reduced variant does not change the answers, because its weights are a modification of the base model’s weights.\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>Which license governs?\u003C\u002Fstrong> Find the \u003Ccode>base_model\u003C\u002Fcode> on the variant’s card and open the base repository’s license. Do not rely on the variant’s license tag alone; \u003Ca href=\"\u002Flearn\u002Fllm-model-licenses\u002F\">our license comparison\u003C\u002Fa> shows a case where the two disagree.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Do you need approval to download?\u003C\u002Fstrong> Llama 3.3 and Gemma 3 are gated on Hugging Face and need manual approval (checked 14 November 2026).\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Is your use on a restricted list?\u003C\u002Fstrong> MIT and Apache 2.0 name no restricted purposes; the law still applies. For Llama 3.3 and Gemma 3, read the use policy against your actual use case.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Is your company very large?\u003C\u002Fstrong> Only the Llama license has a threshold: more than 700 million monthly active users on the Llama 3.3 release date.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Will you share the weights or a modified copy?\u003C\u002Fstrong> Uploading a quantized file or a fine-tune is redistribution. Each license then has the notice duties listed in the table.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Will you serve the model to other people?\u003C\u002Fstrong> Gemma’s terms treat a hosted service as distribution. Llama’s terms attach their attribution duties to anything you “make available”. The MIT and Apache 2.0 texts do not mention hosting.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Will you publish a new model?\u003C\u002Fstrong> Under Llama 3.3, if you use Llama materials or their outputs to create, train or improve a model that you distribute or make available, its name must start with “Llama”.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Do you want to call it open source?\u003C\u002Fstrong> Then check all three OSAID elements, not only the license on the weights. If the developer has not published training code and data information, the accurate description is “open-weight”, followed by the license name.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>This list describes license texts, not how they apply to your facts. It is not legal advice. For anything with commercial or legal stakes, have a lawyer read the full terms against your situation.\u003C\u002Fp>\n\u003Ch2 id=\"common-misconceptions\">Common misconceptions\u003C\u002Fh2>\n\u003Cp>\u003Cstrong>“If I can download it, it is open source.”\u003C\u002Fstrong> Downloadable weights make a model open-weight. The OSAID also asks for data information and code, and for terms that allow any purpose.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>“Apache 2.0 weights make it open source AI.”\u003C\u002Fstrong> An OSI-approved license on the parameters covers one of three elements. The other two, data information and code, are separate requirements.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>“Open source means no obligations.”\u003C\u002Fstrong> MIT requires the copyright and permission notice in “all copies or substantial portions of the Software”. Apache 2.0 requires the license text, prominent notices on changed files and the original attribution notices when you redistribute.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>“All Gemma models share one license.”\u003C\u002Fstrong> They do not. The Gemma Terms of Use list the models they cover (Gemma 1, 1.1, 2, 3 and 3n) and say “For Gemma 4 terms, see the Gemma 4 license”, which Google publishes as the \u003Ca href=\"https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fdocs\u002Fgemma_4_license\" rel=\"noopener\" class=\"ext\">Apache License 2.0\u003C\u002Fa>. A refusal-reduced \u003Ca href=\"\u002Fmodels\u002Fgemma-4-26b-a4b-uncensored\u002F\">Gemma 4 variant\u003C\u002Fa> and a Gemma 3 variant therefore come with different terms.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>“A refusal-reduced variant comes with a fresh license.”\u003C\u002Fstrong> Editing weights does not remove the base terms. An \u003Ca href=\"\u002Fmodels\u002Fllama-3-3-70b-abliterated\u002F\">abliterated Llama 3.3 70B\u003C\u002Fa> still carries the Llama 3.3 license and its Acceptable Use Policy.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>“Nobody knows if weights are copyrighted, so the license does not bind me.”\u003C\u002Fstrong> OSI’s FAQ says the legal status of parameters “will become clearer over time”. Do not read that uncertainty as permission. Meta’s and Google’s terms both say that using the materials means you agree to be bound by them.\u003C\u002Fp>\n\u003Ch2 id=\"how-we-describe-models-on-this-site\">How we describe models on this site\u003C\u002Fh2>\n\u003Cp>Our \u003Ca href=\"\u002Feditorial-policy\u002F\">editorial policy\u003C\u002Fa> uses “open-weight” whenever the weights can be downloaded, then names the license. We use “open source” only for a model that meets the OSAID, or inside a quote. In practice that gives descriptions like these:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>GLM-5.3-Flash: open-weight, MIT license.\u003C\u002Fli>\n\u003Cli>Gemma 4 26B A4B: open-weight, Apache 2.0.\u003C\u002Fli>\n\u003Cli>Llama 3.3 70B Instruct: open-weight, Llama 3.3 Community License, gated download.\u003C\u002Fli>\n\u003Cli>Gemma 3 27B: open-weight, Gemma Terms of Use, gated download.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>The same discipline applies to the word “uncensored”, which tells you even less than “open-weight”. \u003Ca href=\"\u002Flearn\u002Fwhat-is-uncensored-ai\u002F\">What uncensored AI means\u003C\u002Fa> covers that label.\u003C\u002Fp>\n",[17,21,24,27,30,33,36],{"id":18,"text":19,"level":20},"two-separate-questions","Two separate questions",2,{"id":22,"text":23,"level":20},"what-the-osi-definition-requires","What the OSI definition requires",{"id":25,"text":26,"level":20},"where-open-weight-releases-stop","Where open-weight releases stop",{"id":28,"text":29,"level":20},"what-each-license-lets-you-do-with-the-weights","What each license lets you do with the weights",{"id":31,"text":32,"level":20},"a-decision-list-for-your-use-case","A decision list for your use case",{"id":34,"text":35,"level":20},"common-misconceptions","Common misconceptions",{"id":37,"text":38,"level":20},"how-we-describe-models-on-this-site","How we describe models on this site",1694,[41,45,48,51,54,57,60,63,67,71],{"title":42,"publisher":43,"url":44},"The Open Source AI Definition 1.0","Open Source Initiative","https:\u002F\u002Fopensource.org\u002Fai\u002Fopen-source-ai-definition",{"title":46,"publisher":43,"url":47},"Open Source AI Definition: FAQ","https:\u002F\u002Fopensource.org\u002Fai\u002Ffaq",{"title":49,"publisher":43,"url":50},"Open Weights: not quite what you've been told","https:\u002F\u002Fopensource.org\u002Fai\u002Fopen-weights",{"title":52,"publisher":43,"url":53},"Meta's LLaMa license is not Open Source","https:\u002F\u002Fopensource.org\u002Fblog\u002Fmetas-llama-2-license-is-not-open-source",{"title":55,"publisher":43,"url":56},"OSI Approved Licenses","https:\u002F\u002Fopensource.org\u002Flicenses",{"title":58,"publisher":43,"url":59},"The MIT License","https:\u002F\u002Fopensource.org\u002Flicense\u002Fmit",{"title":61,"publisher":43,"url":62},"Apache License, Version 2.0","https:\u002F\u002Fopensource.org\u002Flicense\u002Fapache-2-0",{"title":64,"publisher":65,"url":66},"Llama 3.3 Community License Agreement","Meta","https:\u002F\u002Fwww.llama.com\u002Fllama3_3\u002Flicense\u002F",{"title":68,"publisher":69,"url":70},"Gemma Terms of Use","Google AI for Developers","https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms",{"title":72,"publisher":69,"url":73},"Gemma 4 license (Apache License 2.0)","https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fdocs\u002Fgemma_4_license",[75,78,81,84],{"q":76,"a":77},"Is Llama 3.3 open source?","Not by the OSI's standard. Its license restricts uses through an Acceptable Use Policy and requires very large companies to ask Meta for a separate license, and OSI rejected the same two features in the Llama 2 license.",{"q":79,"a":80},"Is a model released under Apache 2.0 open source AI?","Not automatically. Apache 2.0 can cover the parameters, but the OSI definition also requires the training code and enough information about the training data for a skilled person to build a substantially equivalent system.",{"q":82,"a":83},"Does the OSI definition require publishing the training data?","No. It requires detailed data information and a listing of public and third-party data sources. Data that cannot be shared, such as medical records, may be described instead of released.",{"q":85,"a":86},"Does a refusal-reduced variant get its own license?","No. Its weights are a modification of the base model, so the base license still governs what the variant may be used for, whatever license tag the variant's card shows.",[88,92,96,100],{"path":89,"title":90,"description":91},"\u002Flearn\u002Fllm-model-licenses\u002F","Open-weight model licenses compared","Commercial use, notices, naming rules, user thresholds and use policies under MIT, Apache 2.0, Llama 3.3, Llama 3.1 and Gemma terms, and what variants inherit.",{"path":93,"title":94,"description":95},"\u002Flearn\u002Fwhat-is-uncensored-ai\u002F","What is uncensored AI?","Uncensored AI usually means an open-weight model changed to decline fewer requests. The main types, what the label does not promise and who stays responsible.",{"path":97,"title":98,"description":99},"\u002Flearn\u002Flocal-vs-cloud-llm\u002F","Local vs hosted uncensored LLMs","Run a refusal-reduced model yourself or call a hosted API? Privacy, latency, upkeep and a worked cost comparison in which every input is a stated assumption.",{"path":101,"title":102,"description":103},"\u002Fmodels\u002F","Model catalogue","Every catalogued model with sources, filters and sorting.",null,[106,89,107,108,109,93],"\u002Flearn\u002Fwhat-is-abliteration\u002F","\u002Fmodels\u002Fgemma-4-26b-a4b-uncensored\u002F","\u002Fmodels\u002Fllama-3-3-70b-abliterated\u002F","\u002Feditorial-policy\u002F",1794622336927]