Gemma 4 is here, and now Apache 2.0

The license is the news

On April 2, 2026, Google released Gemma 4, and the headline isn't a benchmark. It's the license. Gemma 4 ships under Apache 2.0, replacing the custom "Gemma Terms of Use" that governed every earlier version. For anyone who wants to own the model they run, that one change matters more than any leaderboard.

Why a license is the story

Earlier Gemma releases were open weights, but they came with strings: a bespoke license carrying use restrictions Google could revise, plus an acceptable-use policy you had to pass downstream. Apache 2.0 cuts the strings. It's a permissive, decades-tested open-source license: no field-of-use limits, no model-specific clauses, commercial use explicitly fine.

You can fine-tune it, redistribute it, embed it in a product, and run it on an air-gapped machine that never phones home. And nobody can rewrite the terms out from under you after the fact. That last part is the quiet revolution: a weight file under Apache 2.0 is a thing you possess, not a service you're licensed to access.

What you're actually getting

Gemma 4 is not a consolation model. The family spans:

• E2B / E4B: compact edge variants built to run on phones and laptops.
• 26B MoE: a mixture-of-experts model with roughly 3.8B active parameters, so it punches well above its compute cost.
• 31B Dense: the heavyweight for serious local reasoning.

Across the line: 128K–256K context windows and native multimodality. Image and video go in, with audio understanding on the E2B/E4B edge variants. All of it runs offline on consumer GPUs and Apple Silicon. No API key, no per-token meter, no request leaving your network.

The shift this confirms

A year ago, "run it locally" meant accepting a real quality tax. That gap has closed for the work most teams actually do. For grounded, document-level question answering, a well-served Gemma 4 on your own hardware is no longer the compromise option. It's the sensible default. The bottleneck stopped being model size some time ago; what matters now is retrieval quality and how cleanly you feed the model its context.

Be precise about what changed. The weights were already downloadable; what Apache 2.0 removes is the permission layer sitting on top of them. Before, "open" came with an asterisk: a license a vendor wrote and could amend. Now the asterisk is gone. You're not asking anyone's blessing to put a state-of-the-art model on a box in a locked room and keep it there.

For a regulated team, that distinction is the whole ballgame. A custom license with revisable use terms is a dependency you can't fully audit, because its meaning can move. A standard open-source license is a known quantity: the same one already running half the infrastructure you trust. Sovereignty isn't only about where the bytes sit; it's about whether the rules governing them can change without you.

That's the foundation pdf2okf is built on. We don't want you renting a model through someone else's terms of use. We want you owning the model, the hardware, and the answer it gives. Gemma 4 under Apache 2.0 is exactly the kind of model pdf2okf runs against locally, or that you point it at with your own key. Own your model. It's the first of three things worth owning: the format you keep your knowledge in, and the stack you build on, are the other two.