Google Drops the Custom License: Gemma 4 Goes Apache 2.0

Google’s been on a roll with Gemini lately, but if you want to actually run something on your own hardware instead of renting it through their API, you’ve been stuck waiting. The last open-weight model from them, Gemma 3, came out over a year ago, which is basically ancient history in this space. Today they’re finally shipping Gemma 4, and there’s a pleasant surprise: they ditched the custom Gemma license.

I’ve been annoyed by proprietary AI licenses for a while now. They always have some weird clause that makes you wonder if you’re accidentally violating terms by using the model in a certain way. Google seems to have heard that feedback loud and clear. Gemma 4 is under Apache 2.0, which is about as permissive as it gets. You can modify it, distribute it, sell it, whatever. No strings attached.

What’s actually shipping? Four model sizes, all designed for local usage. The two big ones are a 26B Mixture of Experts variant and a 31B Dense model. Google claims the 26B MoE can run unquantized in bfloat16 on a single 80GB H100 GPU. That’s a $20,000 card, sure, but it’s still local hardware you can own. If you quantize down to lower precision, both models should fit on consumer GPUs. I’m skeptical about how well they’ll run on something like an RTX 4090, but we’ll see.

The latency focus is interesting. The 26B MoE only activates 3.8 billion of its 26 billion parameters during inference, which means it’s fast. Like, really fast for its size. Google’s claiming significantly higher tokens-per-second than comparable models. The 31B Dense is the opposite play — more about quality than speed, but they expect developers to fine-tune it for specific use cases. That makes sense if you need something specialized.

I’m not sure how these stack up against Llama 4 or Qwen 2.5 yet. Benchmarks will tell the real story. But the licensing change alone is a big win. More permissive licenses mean more experimentation, more fine-tuning, more weird edge case deployments. And that’s good for everyone building on top of this stuff.