Google just released Gemma 4 under Apache 2.0 licensing. Google also has a track record of killing products faster than most people change their passwords. See the tension here?
The company announced four new open models in 2026, ranging from 2 billion to 31 billion parameters. They’re built on the same technology powering Gemini 3, which means they’re actually capable models, not just PR stunts dressed up as open source contributions.
What Apache 2.0 Actually Means
This is the first time Google’s used Apache 2.0 for their Gemma line. Previous versions came with Google’s own licensing terms, which made some developers nervous about building commercial products on top of them. Apache 2.0 is the real deal—permissive, well-understood, and actually open.
You can use these models commercially. You can modify them. You can redistribute them. You don’t need to ask Google for permission or worry about surprise license changes down the road. That’s a significant shift from their previous approach.
Four Models, Four Use Cases
The size range matters more than it sounds. A 2 billion parameter model runs on hardware most developers actually own. A 31 billion parameter model needs serious infrastructure but delivers serious performance. Google’s covering the spectrum from “I want to experiment on my laptop” to “I’m building production systems.”
These models support multimodal capabilities, meaning they handle more than just text. Images, potentially other data types—the usual modern AI feature set. Nothing shocking here, but nothing disappointing either.
The Cynical Take
Let’s be honest about what’s happening. Google watched Meta release Llama models under permissive licenses and saw the entire open source AI community rally around them. They watched Mistral and others eat into mindshare. They realized their previous licensing approach was costing them developer goodwill.
This release is Google playing catch-up in the open source space. They’re late to the Apache 2.0 party, and they know it.
But you know what? Being late doesn’t make the models bad. It just makes Google’s motivations transparent. They want developers building on Gemma instead of Llama. They want researchers citing Gemma papers. They want to be part of the open source conversation instead of watching from the sidelines.
The Practical Take
If you’re a developer, you now have another solid option for open models. The 2B model is worth testing for edge deployment scenarios. The 31B model competes with other large open models in that weight class. The middle sizes give you optimization options.
The Gemini 3 technology foundation means these models benefit from Google’s research budget and infrastructure. That’s not nothing. They’ve been training models at scale longer than most companies have existed.
For researchers, Apache 2.0 licensing removes barriers to experimentation and publication. You can fine-tune these models, study their behavior, and share your findings without legal concerns.
What This Changes
The open model space just got more competitive. Meta, Mistral, Google—they’re all offering capable models under permissive licenses now. That competition benefits everyone building AI applications.
Prices for API access will continue dropping as these open alternatives improve. Vendor lock-in becomes harder to justify when you can run comparable models yourself. The power dynamic shifts toward developers and away from AI providers.
Google’s also signaling they’re serious about open source AI, at least for now. Whether they maintain this commitment long-term depends on how well it serves their business interests. But right now, in 2026, they’re putting real models with real capabilities into the open.
Should You Care?
If you’re building AI applications, yes. More options mean better use in negotiations with API providers and more flexibility in your architecture. If you’re researching AI, absolutely. Apache 2.0 licensing removes friction from your work.
If you’re just watching the AI space evolve, this is another data point showing that open source AI is becoming the norm, not the exception. The major players are competing on openness now, which seemed unlikely just a few years ago.
Google released capable models under a truly open license. That’s good for developers, good for researchers, and good for the AI ecosystem. Whether Google maintains this approach long-term is anyone’s guess, but today’s release is worth paying attention to.
đź•’ Published: