{"course":{"slug":"running-models-yourself","title":"Running Models Yourself","url":"https://addaly.com/learn/running-models-yourself","level":"intermediate","track":"deep"},"lesson":{"slug":"the-open-weights-landscape","title":"What \"open\" means, and what it does not","url":"https://addaly.com/learn/running-models-yourself/the-open-weights-landscape","position":2,"of":82,"minutes":10,"answer":"\"Open\" splits into open weights, open source and open data: most famous models are only the first, and a base model's licence and use policy follow every fine-tune and quant made from it.","keyIdea":"\"Open\" splits into open weights, open source and open data: most famous models are only the first, and a base model's licence and use policy follow every fine-tune and quant made from it."},"body":{"format":"markdown","text":"## Three different things get called open\n\nThe word is doing too much work. Separate it into three claims, because only one of them is usually true.\n\n**Open weights.** The trained parameters are downloadable. You can run them, quantise them, fine-tune them, and inspect every number. This is the common case and it is what this course is about.\n\n**Open source.** The licence meets the Open Source Definition: no restriction on field of use, no restriction on who may use it, no conditions on downstream users. Apache-2.0 and MIT models qualify. Most of the famous ones do not.\n\n**Open data.** The training corpus is published, so the model can be reproduced from scratch. This is rare. AI2's OLMo and the Hugging Face SmolLM family publish theirs; almost nobody else does. Without it you cannot audit what the model was trained on, which matters for contamination, for bias claims, and for any question about copyright.\n\nA model can be open weights and closed source and closed data at the same time, and most are.\n\n## The licences you will meet\n\n- **Apache-2.0 / MIT.** Genuinely permissive. Commercial use, redistribution, modification, no user cap. Most current Qwen releases (in Qwen2.5, the 3B is under a research-only licence and the 72B under Qwen's own licence with a 100-million-user clause), most of Mistral's small releases (Ministral 8B is under the Mistral Research Licence), Falcon-7B and Falcon-40B (later Falcon releases use TII's own Falcon licence), OLMo, DeepSeek's releases since R1, Phi. Check each model, not the family. If you want to stop thinking about licences, stay here.\n- **Llama Community Licence.** Meta's own terms, not open source. Commercial use is allowed, but there is a monthly-active-user threshold above which you must ask Meta for a separate licence, an acceptable use policy attached, and a naming requirement: derivative models must carry \"Llama\" in the name and the attribution notice must be reproduced.\n- **Gemma Terms of Use.** Google's. Commercial use allowed, no user cap, but a use policy that binds you *and* everyone you distribute to, and Google reserves the right to require you to stop using a model remotely if it is being used in breach.\n- **Non-commercial licences** (CC-BY-NC and friends). Research only. Several strong models sit here and people deploy them anyway; that is an infringement, not a grey area.\n- **Gated but permissive.** Some Apache-2.0 models still require you to accept terms on Hugging Face and be approved before download. The gate is a distribution control, not a licence term.\n\n## The part almost nobody checks\n\nThe base model's licence follows its descendants. A fine-tune of a Llama model is still under the Llama licence no matter what the uploader wrote in their own README. When you download `SomeLab/CoolModel-8B`, the questions are: what was it fine-tuned from, and what did *that* licence say. The model card usually names the base model in the first paragraph; the config file names the architecture, which is a strong hint.\n\nThe same applies to synthetic training data. If a dataset was generated by a commercial API, that API's terms may forbid using it to train a competing model. Several popular fine-tunes carry that history and do not mention it.\n\n## Where the law is unsettled\n\nTwo questions have no settled answer, and the answer differs by country:\n\n1. **Was training on copyrighted text lawful?** The United States is arguing this through fair use, the EU through the text-and-data-mining exception and its opt-out, Japan has a broad exception, the United Kingdom has consulted repeatedly without concluding. Cases are live in several jurisdictions and have gone both ways on different facts.\n2. **Who owns the output?** Several offices, including the US Copyright Office, have said purely machine-generated output is not protectable, while human-authored contributions around it may be. What counts as enough human contribution is not defined.\n\nThis course does not resolve either, because they are not resolved. What it can tell you is the practical shape: running a model locally does not change your copyright position at all, and a licence from a model provider is a licence to use the weights, never a warranty that the weights were lawfully made. If your use is commercially serious, that is a question for a lawyer in your jurisdiction, not for a forum.\n\n## The practical routine\n\nBefore downloading anything you intend to build on:\n\n1. Open the model card and find the licence field and the base model.\n2. Read the acceptable use policy if there is one. It is short and it binds you.\n3. Save a copy of the licence text alongside the weights. Licences have been changed after release; the copy you agreed to is the one you can point at.\n4. If you will redistribute anything (a fine-tune, a quant, a container image) check the attribution and naming clauses, because those are the ones that catch people.\n\nThat takes ten minutes and removes an entire class of problem that only ever surfaces once something you built has become important."},"check":{"question":"You fine-tune a Llama-family model on your own data and want to publish the result on Hugging Face under Apache-2.0, since your training code and your dataset are both yours. What is wrong with that plan?","options":[{"text":"Nothing, provided you credit Meta in the README: attribution satisfies the community licence for derivative works.","correct":false,"tempting":"Reduces a licence to a courtesy; the naming and downstream-terms clauses are obligations, not credits, and cannot be swapped for a mention."},{"text":"Fine-tuned weights cannot be published at all, because the resulting parameters are a derivative of copyrighted training data.","correct":false,"tempting":"Imports an unsettled copyright question as if it were settled law, and would prohibit the entire open fine-tune ecosystem."},{"text":"The derivative weights remain under the original community licence, which you cannot relicense, and it carries naming and use-policy conditions that pass to your users.","correct":true},{"text":"Apache-2.0 is the wrong choice only because it lacks a patent grant appropriate for model weights.","correct":false,"tempting":"Invents a technical objection; Apache-2.0 does contain a patent grant, and the actual problem is that you never had the right to choose the licence."}]},"prev":{"slug":"why-run-a-model-yourself","title":"Why run a model yourself, and when not to","url":"https://addaly.com/learn/running-models-yourself/why-run-a-model-yourself"},"next":{"slug":"reading-a-model-card","title":"Reading a model card without being sold to","url":"https://addaly.com/learn/running-models-yourself/reading-a-model-card"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'What \"open\" means, and what it does not', in Running Models Yourself. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/running-models-yourself/the-open-weights-landscape (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What “open” means, and what it does not","courseTitle":"Running Models Yourself","module":{"index":1,"slug":"deciding-to-run-it-yourself","title":"Deciding to run it yourself"},"answer":null,"keyIdea":"“Open” splits into open weights, open source and open data: most famous models are only the first, and a base model’s licence and use policy follow every fine-tune and quant made from it.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"three-different-things-get-called-open","runs":[{"text":"Three different things get called open"}]},{"t":"p","runs":[{"text":"The word is doing too much work. Separate it into three claims, because only one of them is usually true."}]},{"t":"p","runs":[{"text":"Open weights.","b":true},{"text":" The trained parameters are downloadable. You can run them, quantise them, fine-tune them, and inspect every number. This is the common case and it is what this course is about."}]},{"t":"p","runs":[{"text":"Open source.","b":true},{"text":" The licence meets the Open Source Definition: no restriction on field of use, no restriction on who may use it, no conditions on downstream users. Apache-2.0 and MIT models qualify. Most of the famous ones do not."}]},{"t":"p","runs":[{"text":"Open data.","b":true},{"text":" The training corpus is published, so the model can be reproduced from scratch. This is rare. AI2’s OLMo and the Hugging Face SmolLM family publish theirs; almost nobody else does. Without it you cannot audit what the model was trained on, which matters for contamination, for bias claims, and for any question about copyright."}]},{"t":"p","runs":[{"text":"A model can be open weights and closed source and closed data at the same time, and most are."}]},{"t":"h2","id":"the-licences-you-will-meet","runs":[{"text":"The licences you will meet"}]},{"t":"ul","items":[[{"text":"Apache-2.0 / MIT.","b":true},{"text":" Genuinely permissive. Commercial use, redistribution, modification, no user cap. Most current Qwen releases (in Qwen2.5, the 3B is under a research-only licence and the 72B under Qwen’s own licence with a 100-million-user clause), most of Mistral’s small releases (Ministral 8B is under the Mistral Research Licence), Falcon-7B and Falcon-40B (later Falcon releases use TII’s own Falcon licence), OLMo, DeepSeek’s releases since R1, Phi. Check each model, not the family. If you want to stop thinking about licences, stay here."}],[{"text":"Llama Community Licence.","b":true},{"text":" Meta’s own terms, not open source. Commercial use is allowed, but there is a monthly-active-user threshold above which you must ask Meta for a separate licence, an acceptable use policy attached, and a naming requirement: derivative models must carry “Llama” in the name and the attribution notice must be reproduced."}],[{"text":"Gemma Terms of Use.","b":true},{"text":" Google’s. Commercial use allowed, no user cap, but a use policy that binds you "},{"text":"and","i":true},{"text":" everyone you distribute to, and Google reserves the right to require you to stop using a model remotely if it is being used in breach."}],[{"text":"Non-commercial licences","b":true},{"text":" (CC-BY-NC and friends). Research only. Several strong models sit here and people deploy them anyway; that is an infringement, not a grey area."}],[{"text":"Gated but permissive.","b":true},{"text":" Some Apache-2.0 models still require you to accept terms on Hugging Face and be approved before download. The gate is a distribution control, not a licence term."}]]},{"t":"h2","id":"the-part-almost-nobody-checks","runs":[{"text":"The part almost nobody checks"}]},{"t":"p","runs":[{"text":"The base model’s licence follows its descendants. A fine-tune of a Llama model is still under the Llama licence no matter what the uploader wrote in their own README. When you download "},{"text":"SomeLab/CoolModel-8B","code":true},{"text":", the questions are: what was it fine-tuned from, and what did "},{"text":"that","i":true},{"text":" licence say. The model card usually names the base model in the first paragraph; the config file names the architecture, which is a strong hint."}]},{"t":"p","runs":[{"text":"The same applies to synthetic training data. If a dataset was generated by a commercial API, that API’s terms may forbid using it to train a competing model. Several popular fine-tunes carry that history and do not mention it."}]},{"t":"h2","id":"where-the-law-is-unsettled","runs":[{"text":"Where the law is unsettled"}]},{"t":"p","runs":[{"text":"Two questions have no settled answer, and the answer differs by country:"}]},{"t":"ol","start":1,"items":[[{"text":"Was training on copyrighted text lawful?","b":true},{"text":" The United States is arguing this through fair use, the EU through the text-and-data-mining exception and its opt-out, Japan has a broad exception, the United Kingdom has consulted repeatedly without concluding. Cases are live in several jurisdictions and have gone both ways on different facts."}],[{"text":"Who owns the output?","b":true},{"text":" Several offices, including the US Copyright Office, have said purely machine-generated output is not protectable, while human-authored contributions around it may be. What counts as enough human contribution is not defined."}]]},{"t":"p","runs":[{"text":"This course does not resolve either, because they are not resolved. What it can tell you is the practical shape: running a model locally does not change your copyright position at all, and a licence from a model provider is a licence to use the weights, never a warranty that the weights were lawfully made. If your use is commercially serious, that is a question for a lawyer in your jurisdiction, not for a forum."}]},{"t":"h2","id":"the-practical-routine","runs":[{"text":"The practical routine"}]},{"t":"p","runs":[{"text":"Before downloading anything you intend to build on:"}]},{"t":"ol","start":1,"items":[[{"text":"Open the model card and find the licence field and the base model."}],[{"text":"Read the acceptable use policy if there is one. It is short and it binds you."}],[{"text":"Save a copy of the licence text alongside the weights. Licences have been changed after release; the copy you agreed to is the one you can point at."}],[{"text":"If you will redistribute anything (a fine-tune, a quant, a container image) check the attribution and naming clauses, because those are the ones that catch people."}]]},{"t":"p","runs":[{"text":"That takes ten minutes and removes an entire class of problem that only ever surfaces once something you built has become important."}]}],"check":{"question":[{"text":"You fine-tune a Llama-family model on your own data and want to publish the result on Hugging Face under Apache-2.0, since your training code and your dataset are both yours. What is wrong with that plan?"}],"options":[{"text":[{"text":"Nothing, provided you credit Meta in the README: attribution satisfies the community licence for derivative works."}],"correct":false,"tempting":[{"text":"Reduces a licence to a courtesy; the naming and downstream-terms clauses are obligations, not credits, and cannot be swapped for a mention."}]},{"text":[{"text":"Fine-tuned weights cannot be published at all, because the resulting parameters are a derivative of copyrighted training data."}],"correct":false,"tempting":[{"text":"Imports an unsettled copyright question as if it were settled law, and would prohibit the entire open fine-tune ecosystem."}]},{"text":[{"text":"The derivative weights remain under the original community licence, which you cannot relicense, and it carries naming and use-policy conditions that pass to your users."}],"correct":true,"tempting":null},{"text":[{"text":"Apache-2.0 is the wrong choice only because it lacks a patent grant appropriate for model weights."}],"correct":false,"tempting":[{"text":"Invents a technical objection; Apache-2.0 does contain a patent grant, and the actual problem is that you never had the right to choose the licence."}]}],"order":[1,0,3,2],"why":null},"prev":{"slug":"why-run-a-model-yourself","title":"Why run a model yourself, and when not to","minutes":6},"next":{"slug":"reading-a-model-card","title":"Reading a model card without being sold to","minutes":9}}}