Gemma licence vs Apache for specialist adaptation

Who this page is for

Open this page when the proposed base is a Gemma revision and the legal question is whether Google’s Terms of Use are interchangeable with an Apache file you already approved on Qwen or Phi. The people who need the answer are counsel comparing two publisher documents, a field-IT owner who wants a 2B-class image on a locked tablet, and a specialist who will still confirm every extract. We will not teach a Gemma notebook here, and we will not publish a size-versus-quality league.

If the brief only said “get a small model”, start at the commercial SLM page. If the brief only said “we need Apache on the weights”, start at Qwen or the current Mistral 3 announcement and return only if a Gemma revision is still the intended artefact. If the brief said “open source” and meant OSAID, start at OLMo. Gemma’s job on this site is the Terms-versus-later-Apache split, and the fact that a fine-tune does not retire the Terms.

Gemma Terms are a Google document, not a missing Apache stamp

Google publishes Gemma under Gemma Terms of Use, sitting next to an acceptable-use policy. That is a different bargain from Apache-2.0 on a Qwen3 LICENSE file or MIT on a cited Phi card. You can often download, run and adapt Gemma. You do not get a sentence you can file as “it is open source, ship the merge”. The citation on this page is Google’s published Terms, not a blog that paraphrased them into Apache.

Procurement that already signed Apache-2.0 Qwen or MIT Phi has not signed Gemma. Put the Terms URL next to the other family’s file and ask the same questions once: commercial use, redistribution of the adapter or merge, acceptable use, and what must appear if a product embeds the materials. Those answers come from counsel reading Google’s text. They do not come from a slide that says “Gemma is the small open one”.

Apache-2.0 is simpler for teams that want notice-and-disclaimer and nothing else. Gemma Terms are not trying to be that story. Treating them as a formality because the weights are downloadable is how a later diligence review finds a second publisher document nobody filed. Downloadable parameters are open weights. The Terms are the grant you actually have on an earlier revision.

The Terms survive a fine-tune

A LoRA rank change, a merged single-task image, and a quantised tablet build are still Gemma materials if you started from Gemma weights. The adapter is not a new legal species. Internal-only serving can change who sees a product page. It does not automatically delete Google’s document. If you distribute the adapter, ship a product that uses the materials, or hand the merge to a vendor, have counsel read the current Terms for that distribution story before anyone calls the artefact “ours”.

Full-parameter updates do not buy a new licence either. Touching every weight is a training choice. It is not a prestige upgrade that converts Terms into Apache. Teams who “just want to own the model” after a full fine-tune are mixing a technical story with a publisher story. You own the work you did. You still started from Gemma. Write both facts into the handover.

This page will not reprint Meta’s naming or attribution clauses, and it will not invent a Gemma equivalent. If you need that other bargain, read /models/llama. If you need notice-only Apache on a named Qwen3 revision, read /models/qwen. Gemma’s surviving duty is whatever the Terms and acceptable-use policy say for the revision you actually froze. Cite that revision’s URL and the date you read it.

• Treat a distributed adapter or merge as a derivative of the Gemma materials until counsel says otherwise.
• Keep the Terms (or the later licence file on that revision) with the artefact.
• Write the acceptable-use sentence the workflow actually matches.
• Do not file a full fine-tune as an escape from Google’s document.

A later Apache-labelled revision is a different artefact

Model families drift. Public write-ups in 2026 describe some newer Gemma revisions as Apache-2.0. We will not treat that as true for every Gemma weight, and we will not back-apply an Apache line to an earlier Terms checkpoint. The only safe sentence in a statement of work is the licence line on the revision you will train, with a URL and a date. If that line still points at Gemma Terms, the fine-tune conversation is a Terms conversation.

If that line names Apache-2.0, the grant for that artefact may be the Apache file — for that hash, not for the family. You still do not inherit Apache for a 2B image you downloaded last year. You still do not inherit OSAID. Apache on parameters is an OSI-approved software licence for those files. The Open Source Initiative’s AI definition also asks for the preferred form to modify the system, including data information and the code needed to derive the parameters. Gemma, even on an Apache-labelled later revision, is usually open weights with a software licence, not a reproduced stack.

Do not let a vendor say “Gemma is Apache now” as a reason to skip the URL. Do not let an engineer say “we fine-tuned so the old Terms no longer matter” as a reason to skip the URL. Two revisions can sit in the same object store under the same family name. Label them. The later Apache-labelled file does not wash the earlier Terms file.

Acceptable use is in the same pack as the Terms

Google’s acceptable-use policy sits beside the Terms. A workflow the policy rules out is not rescued by a cleaner dataset, a smaller rank, or a locked device. Write the intended use in one sentence a specialist would recognise: suggest agreed fields from a photo or PDF a handler still confirms; fill catalogue attributes a merchandiser still approves. If that sentence is open-ended advice to customers, or a decision the firm is not allowed to automate, Gemma is the wrong conversation until the workflow changes.

Acceptable use is not a substitute for your own conduct rules. Insurance correspondence a handler must still send, and retail copy a merchandiser must still approve, are the usual shape. A model that “decides coverage” on a tablet is not a Terms problem first. It is a brief problem. The device does not create authority the desk did not have.

Size bands change the box, not the document you signed

Compact 2B-class Gemma cards enter on-device and private-footprint conversations because they fit. That is a hosting reason. It is not evidence they will extract your policy fields, and it is not a licence discount. 9B–27B-class cards are the mid band when the team already operates that memory plan. The Terms (or a later licence line on that specific revision) still apply across those bands. Size is not a quality ranking on this site.

Pick the band from hardware you actually operate — including the tablet, the store box, or the private GPU — and from a written quality bar on unused test documents. Do not pick 2B because a slide said small models are cheaper. Do not pick 27B because a listicle said it won a public exam. A compact Gemma that misses the bar is unfinished work. Whether you should shrink at all still lives on the commercial SLM page. This page is the publisher document that rides with whichever band you picked.

On-device is a hosting sentence, not independence

A locked field tablet running a 2B-class Gemma image can be the right box for a sparse network, a store floor, or a yard where a laptop is a joke. It is still a box someone updates, someone encrypts, and someone can lose. The Terms still apply to the weights on that flash. Logs can still leave if the app phones home. A fine-tune packaged as a device image does not create sovereignty. It creates a smaller serving plan with the same publisher document attached.

Write the update channel. Write where prompts sit when the device is online. Write who may load a new adapter onto the fleet. Those sentences are the product. A demo that runs offline for five minutes is not an operating plan. The inference page is the environment conversation. This page is why “it runs on our tablet” does not retire Google’s Terms or turn the image into an Apache exhibit.

What changes in the run

A typical Gemma specialist run freezes the named revision and trains adapters. The artefact you store is the adapter or a merged copy, the exact identifier, the Terms text or the later licence file that revision actually publishes, the acceptable-use sentence, and notes from unused test documents. The optimiser does not rewrite Google’s document. A merge into a tablet image does not rewrite it either. Inference location changes only if you actually move the box.

If the named revision’s licence line is still Gemma Terms, the run produces a Terms-bound artefact. If the named revision’s licence line is Apache-2.0, the run produces an Apache-bound artefact for that hash — and you still write the URL so a later operator does not apply that file to last year’s Terms image. If someone later wants a student in another family, read both grants before traces exist. This page will not reprint a community “do not improve another LLM” clause; that fight lives on /models/llama. Gemma’s fight is whether you thought a fine-tune retired the Terms.

Device packaging is part of the run when the brief is on-device extract. You are not done when the trainer exits. You still owe a signed image, a pin of the base hash, the Terms or Apache file inside the image’s notice directory, and a fail-closed path when the specialist is unavailable. A run that only dumps adapter files on a laptop has not met a field brief.

Insurance: on-device field extract

Retail: small-footprint attributes versus a Qwen3 Apache alternative

When Gemma is the wrong first base

Pick another family when procurement requires an OSI-approved licence on the weights and the revision in front of you still names Gemma Terms. Pick another family when counsel asked for OSAID and will not accept open weights plus a Google document. Pick another family when the organisation already standardised on Qwen operations and the only reason for Gemma is a tablet brochure. Those are constraint problems. They are not solved by a cleaner LoRA rank.

Gemma is also the wrong first conversation when the team will not operate a device fleet or a private box this year and the only available path is a hosted endpoint that happens to say Gemma on the SKU. Be honest. A downloadable Terms checkpoint sitting in a zip on a laptop is not a field plan. The inference page is the operating conversation. The closed-API comparison is the prompt-path conversation.

Redistribution is a Terms question even when the box never leaves the building

Handing a merged Gemma image to an outsourcer, a hardware vendor, or a sister company is distribution in the ordinary commercial sense even if the tablet never sees the public internet. Write who receives the file. Write whether they may train further. Write whether they may ship it inside their own product. Those sentences belong in the brief before the first epoch. This page will not invent Google’s answers. It will refuse to treat “we only use it internally” as true when the zip has already gone to a vendor.

Quantised packs from a third-party repository are a second distribution story. A GGUF that dropped the Terms file is not a cleaner artefact. It is a missing exhibit. If you must quantise for the tablet, keep the publisher document in the same directory as the weights and the adapter. Operators who “just grabbed a community quant” often grab a licence hole at the same time.

What handover must include

A useful Gemma handover is the adapter or merged image, the exact base revision, the Terms text or the later licence file that revision publishes, the acceptable-use sentence, unused-test notes from the field or catalogue workflow, and the environment — tablet fleet, store box, or private GPU. It is not a claim that the fine-tune is open-source AI, that a device image is now sovereign, or that a public exam transferred to your book.

Operators still need the base weights and a path that allows the stack. Merging simplifies a single-task tablet image. Keeping adapters separate helps more than one product on one frozen Gemma base if the device can load them. Choose from the operating plan. The LoRA page is the artefact conversation. This page is the publisher document that rides with the artefact — and the warning that a later Apache-labelled hash is not a wash for an earlier Terms image.

Questions that belong in the brief once, in writing

Which Gemma revision will we freeze? Does that revision’s licence line still name Gemma Terms, or does it name Apache-2.0 or something else? May we commercially adapt it? May we redistribute the adapter or the tablet image? Does acceptable use allow this extract or attribute job? Where do prompts sit when the device is online? Did anyone mean OSAID when they said open source? If procurement will only sign Apache, why is Gemma still in the sentence?

Those questions are more useful than a debate about whether Google is “more open this year”. The openness comparison exists so you can point at OSI’s definition without repeating it here. The Qwen page exists so the Apache alternative has its own URL. This page exists so Gemma Terms, a later Apache-labelled hash, and a fine-tune that does not retire either of them have one place to live.

Sources

1. Google, Gemma Terms of Use (Accessed September 2026). Gemma distributions are governed by Google’s Gemma terms and acceptable-use policy unless a specific checkpoint card names a different licence.
2. Open Source Initiative, Open Source AI Definition 1.0 (Accessed September 2026). An open-source AI system must provide the preferred form for modification, including parameters under OSI-approved terms plus the data information and code needed to derive them.
3. Open Source Initiative, Open Weights (Accessed September 2026). Open weights are the published parameters of a trained network. Weights alone are not the preferred form for modifying the system.

Questions we hear first