Gemma 4 opens doors with Apache 2.0, shifting enterprise AI choices.

For two years, enterprise teams evaluating open-weight models had to make a choice that wasn't really a choice. Google's Gemma line delivered the performance, but the license was a hurdle you could trip over. Legal review added weeks. Compliance flagged edge cases. And even when you cleared those hurdles, the terms could change at any time, leaving your deployment on unstable ground. Gemma 4 changes that with a single, decisive move: Apache 2.0. No asterisks. No custom carve-outs. Just the same permissive terms as the rest of the open-weight ecosystem. If you've been holding off on Google's models because of licensing friction, the reason to wait is gone.

What matters most here isn't the benchmark scores, though they're strong. It's that Google has finally aligned its open models with how enterprises actually deploy software. The 26B A4B MoE model, with its 128 small experts and 3.8 billion active parameters, delivers reasoning performance in the 27B, 31B class while running at roughly the speed of a 4B model. That's not a spec sheet curiosity; it's a direct line to lower serving costs and lower latency in production. For teams running coding assistants, document pipelines, or agentic workflows, the MoE variant is the practical pick. And the edge models, with native audio and vision, mean you can keep data on-device for healthcare, field service, or any use case where privacy isn't optional. The architecture choices here aren't just interesting; they're the difference between a model you demo and a model you deploy.

The timing also matters. While some Chinese labs are pulling back from fully open releases, Google is moving the opposite direction, opening up its most capable Gemma yet and explicitly drawing from Gemini 3 research. That's a signal about where the market is heading, and it's a bet on trust. For enterprises, the Apache 2.0 license removes the need for a legal review before you can even start a pilot. It means fine-tuned derivatives can go to production without asking permission. And with serverless deployment on Cloud Run scaling to zero when idle, the cost structure of running open models just changed in your favor. You pay for inference, not idle GPUs.

The practical takeaway is straightforward: if you've been waiting for an open-weight model that combines strong reasoning, native multimodality, function calling, and a license that doesn't require a lawyer, Gemma 4 is the first Google release where you don't have to compromise. The evaluation can now start with the model itself, not with the fine print. That's the shift worth acting on.