Why is Meta's Llama called open-weight rather than open-source?

## Two different claims "Open source" is a legal-category claim about a licence. "Open weight" is a distribution claim about an artefact. Llama satisfies the second and not the first, and conflating them is the single most common error in this area. What Meta actually publishes is the trained parameters, a tokenizer, model cards and inference code, released under a bespoke *Llama Community License Agreement* that you accept before download. Anyone can take those weights, run them on their own hardware with no network call to Meta, inspect activations, fine-tune them, quantise them to run on a laptop, and redistribute the result — subject to the agreement's conditions. ## Why it fails the Open Source Definition The Open Source Definition maintained by the Open Source Initiative sets ten criteria a licence must meet. Llama's agreement breaches at least two of them. **No discrimination against persons or groups.** The additional commercial terms require any licensee whose products exceeded 700 million monthly active users in the calendar month before that version's release to request a licence from Meta, granted at Meta's sole discretion. A licence whose rights depend on who you are is not an open-source licence, however few entities are affected. **No discrimination against fields of endeavour.** The agreement incorporates Meta's Acceptable Use Policy by reference, which bars entire application categories. Open-source licences deliberately refuse to encode such use restrictions — that is why licences with "do no evil" or non-military clauses have never been OSI-approved either. Additionally, the agreement is unilaterally attached to a specific Meta-defined set of materials, imposes attribution and model-naming conditions, and is a different document for each release. None of that is fatal on its own — copyleft licences impose conditions too — but together with the two clauses above it puts Llama firmly outside the category. Separately, the OSI's Open Source AI Definition (published 2024) adds requirements around disclosure of training-data information; Llama does not publish its training corpus, which is another reason the label does not fit even under the AI-specific definition. ## What "open weight" buys you anyway The distinction is not academic pedantry that makes the licence worthless. Open weights give you capabilities a hosted API cannot: - **Data residency and air-gapping.** Inference runs entirely inside your boundary; no prompt leaves your network. This is often the whole reason a regulated organisation chooses Llama. - **No vendor deprecation.** A hosted model can be retired on the vendor's schedule; weights on your disk cannot. You control the upgrade timing. - **Cost structure.** You pay for GPUs and engineers, not per token — which wins at sustained high volume and loses at spiky low volume. - **Modification.** Fine-tuning, quantisation, pruning, and adapter stacking are all available and all permitted by the agreement. - **Auditability of behaviour.** You can reproduce a given output exactly, which matters for regulated review. What open weights do *not* give you is the ability to treat the licence as boilerplate. Because the terms are bespoke, someone has to read them per release. ## Saying it correctly in an interview A weak answer says "Llama is open source, so we can do whatever we like." A strong answer says: the weights are open and commercially usable, the licence is a bespoke Meta agreement rather than an OSI-approved one, the two disqualifying terms are the acceptable-use restrictions and the very-large-operator threshold, and the practical consequences for our team are attribution and naming obligations plus a per-version licence read. That answer demonstrates you would not embarrass the company in a compliance review — which is the actual thing being tested. ## The vocabulary that is now standard The industry has largely settled on a three-way vocabulary: *closed* (weights never leave the vendor, API access only), *open weight* (weights downloadable under a bespoke licence with conditions), and *open source* (weights and code under an OSI-approved licence with no field-of-use restrictions). Llama sits squarely in the middle bucket. Some other open-weight families do ship under standard permissive licences and therefore sit in the third bucket — which is exactly why using the terms precisely matters when comparing options in a design review.