Google released Gemma 4 on April 2, 2026. The press got the headline wrong, in the way the press usually gets these stories wrong: by leading with the benchmarks instead of the license.

Read the license. The model is Apache 2.0. That is the entire story.

Why the license is the story

Most "open" AI model releases in 2026 are not actually open in the sense that anyone who has shipped commercial software would recognize. Llama's various licenses include commercial-use restrictions above a certain user threshold. Custom-license releases like Mistral's commercial-tier models prohibit competitor training. The pattern is "open-ish, with conditions tailored to protect the releasing lab's commercial interests."

Gemma 4 didn't do that. Apache 2.0 is one of the most permissive widely-used open-source licenses. You can fine-tune it for any purpose, run it commercially without notifying anyone, redistribute modifications, and use it as the foundation of a competing commercial product. There are no usage-based clauses. There is no "you can't use this to train a competing model" clause. There is no field-of-use restriction.

For a model built on the same research as Gemini 3.1 Pro — Google's flagship — this is a meaningful concession.

What's in it

The specs that matter for builders:

• Built on the same research foundation as Gemini 3.1 Pro
• Several parameter-size variants (covering edge-device through workstation deployments)
• Multimodal — text plus vision
• Competitive with closed frontier models on most reasoning benchmarks (per Google's own reporting; independent reproduction in progress)
• Apache 2.0 license, no asterisks

The capability gap between Gemma 4 and Gemini 3.1 Pro is real but smaller than you'd expect. Internal reports from teams that have been A/B testing them say the gap is meaningful on the hardest reasoning problems and almost imperceptible on routine work. That maps to the experience of anyone who has been using OSS-friendly models for production: the marginal capability gap matters less than the operational freedom.

Why Google did this

I've thought about this for a few weeks. The most charitable read is the one I think is also the most accurate: Google is making a long-term ecosystem play.

The bet: by making the most permissive frontier-adjacent model in the market, Google becomes the default choice for any developer who wants the option to not be locked into a single inference vendor. Once you build on Gemma 4 for fine-tuning or research or anywhere on-device, you remain a Google ecosystem developer even if your production system is running on a different host.

That's a long game. Anthropic and OpenAI can't easily counter it without releasing their own frontier weights, which they have business-model reasons not to do. Meta could have countered it. Meta chose to ship Muse Spark closed-weight in April instead. So Google has the permissive-frontier lane to themselves.

If Google maintains this stance for two more model generations, the open-weights frontier becomes effectively Google's. Whether that's a good outcome for the ecosystem is a separate question. Right now, it's a builder gift.

What about Apache 2.0 actually buys you

Three things that matter in practice:

One: no audit-the-license overhead. Apache 2.0 is the license your legal team has already approved. You can ship Gemma 4-based products without a custom legal review. That sounds boring. It's worth weeks of timeline.

Two: forks and redistribution. Someone else can fork Gemma 4, fine-tune it on domain data, and redistribute the result under Apache 2.0. That's the foundation of how Hugging Face's specialty-model ecosystem actually works. Gemma 4 is going to spawn hundreds of specialized variants in the next six months.

Three: no rug-pull risk. Google can't retroactively change the terms on already-released Apache 2.0 weights. The weights are out. Whatever Google does in the future, the current Gemma 4 is permanently usable under permissive terms. That's the kind of guarantee that lets you build infrastructure around a model.

Further reading

• Google AI updates — April 2026 — official announcement context
• Hugging Face — Gemma 4 model card — weights + technical specs
• Apache License 2.0 text — read this before assuming you know what it allows