Gemma 4 under Apache 2.0 while Meta retires Llama

Two announcements, eight days apart

On 2 April Google released Gemma 4, five models from a 2.3B effective-parameter edge model to a 31B dense model and a 26B mixture-of-experts with 4B active parameters, each in base and instruction-tuned form. The line in the Hugging Face announcement that we read twice was about the licence. Every Gemma 4 model ships under Apache 2.0. That is the same licence Mistral used for its early releases, and it is a plain OSI-approved permissive licence with no acceptable-use policy attached.

On 8 April Meta Superintelligence Labs released Muse Spark. The weights are closed. Wikipedia's Llama article states that Muse Spark was released as a replacement for Llama, and Meta's own framing positions it as a new model family rather than Llama 5. The lab was formed on 30 June last year, after Zuckerberg's reported dissatisfaction with Llama 4, with Alexandr Wang as chief AI officer and Nat Friedman on product. Llama 4 Scout and Maverick from April 2025 remain the last Llama release.

How we got here

Three years ago the positions were reversed. Meta's LLaMA weights leaked on 3 March 2023 within a week of a gated research release, and by July Meta had turned that accident into a strategy, releasing Llama 2 for commercial use under a custom licence. Llama 3 in April 2024, the 405B model that July, and the multimodal 3.2 release in September made Meta the reference open-weight lab in the United States. The licences were never open source in the OSI sense, with the 700 million monthly user cutoff and the acceptable-use policy, but the weights were there and the ecosystem built on them.

Google was the cautious one. The first Gemma models in February 2024 came with a custom terms-of-use document that included a prohibited-use policy, pass-through obligations on derivatives including distilled models, and a clause reserving Google's right to restrict usage remotely. We wrote at the time that the word open was doing a lot of work in that announcement. That clause is gone now. Under Apache 2.0 there is nothing for Google to restrict.

What Apache 2.0 changes in practice

For most users the day-to-day difference is small, and we do not want to overstate it. People who were going to fine-tune Gemma 3 under its terms will fine-tune Gemma 4 the same way. The difference shows up at the edges. A team distilling Gemma 4 into a smaller model no longer inherits a prohibited-use policy that travels with the derivative. A company legal department that refused custom licences on principle can now say yes. A research group publishing weights derived from Gemma no longer needs to reproduce a use policy in its repository.

The other change is auditability of the licence itself. Custom terms can be revised, and Google's earlier terms said usage could be restricted if Google reasonably believed the agreement was violated. Apache 2.0 is a fixed text with two decades of case history. A model licensed under it today cannot be relicensed out from under you.

The models themselves are competitive enough that the licence matters. The 31B reports 85.2 on MMLU Pro and 84.3 on GPQA Diamond, with 256k context, and the announcement lists day-one support in transformers, llama.cpp, MLX and ONNX. This is not an Apache 2.0 release of a model nobody would use.

What Meta is signalling

We want to be careful here because Meta has said less than the coverage implies. What is on record is that Muse Spark is proprietary, that it comes from a new lab with new leadership, and that it replaces Llama rather than continuing it. Zuckerberg's 2024 argument that open weights were the business case, the Linux analogy, has not been formally withdrawn. It has been left behind by the org chart.

The honest reading is that Meta's open-weight period was a strategy for a lab that was behind, and Meta no longer wants to be the lab that is behind. Whether Muse models will ever be opened is a question the company has not answered as of this week, and we would rather wait for an answer than guess one.

What to watch

The interesting question for the rest of the year is whether Apache 2.0 becomes the default that other labs have to match. Mistral returned to it in early 2025 after experimenting with a custom licence, and several Chinese labs ship under MIT or Apache already. If Google holds this line through Gemma 5, the custom open-weight licence, the thing Meta invented and Google copied in 2024, will have been a two-year detour.

The experiment we would run is boring and useful. Take a Llama 3.3 70B derivative and a Gemma 4 31B derivative and try to get both through the licensing review at three different organisations. Our guess is the Gemma one clears in an afternoon and the Llama one generates a memo. That difference is the whole story of this month, and it is the difference researchers will feel first.