Google releases Gemma 4 under Apache 2.0

Single H100. That's what catches my attention — not the Apache license, not the benchmark charts Google will inevitably wave around. One 80GB GPU running Gemini-class inference.

We've been paying cloud API bills for this quality level for years. Gemini Pro isn't cheap at scale, and now here's the same underlying research baked into something you can self-host. No commercial restrictions means no legal gymnastics for product use.

Look, every 'open' model release comes with an asterisk. Llama 2 restrictions, 'research only' fine print — someone always buries something. Apache 2.0 is actually clean (not MIT-level paranoid-clean, but clean enough). Fine-tune it, wrap it in your own UI, resell it. Nobody's stopping you.

Honestly, this kills the API-only argument for lower-volume internal tools. If you've got dedicated GPU access, the cost math flips completely. We're already benchmarking this against one client project where the API spend was getting uncomfortable.

There's a catch: you still need the H100. That's roughly $2.49/hr on Lambda Labs. But if you're running constant inference load, break-even against Gemini API costs comes sooner than you'd expect.