"The Local Threshold"
Google's Gemma 4 crosses a threshold that matters for what AI becomes in practice, not just in capability benchmarks. A 26-billion parameter mixture-of-experts model that activates only 2 to 4 billion parameters during inference, runs entirely offline on phones and single-board computers, processes vision and audio natively, handles 256K token contexts, and ships under Apache 2.0.
Each of these features has existed separately. Local inference on small models — Llama, Phi, Mistral. Large context windows — Claude, GPT-4. Multimodal input — Gemini, GPT-4V. Permissive licensing — Llama 3. But the combination in a single model that fits on consumer hardware changes the deployment surface. The constraint isn't any one capability but the intersection of capabilities available without an API call.
The licensing shift from restrictive to Apache 2.0 is the quietest and most consequential change. Previous Gemma versions carried use restrictions that limited commercial deployment. Apache 2.0 means anyone can modify, redistribute, and monetize. The model becomes infrastructure — a component that products build on without negotiating terms. This is how open standards win: not by being the best, but by being the most frictionless.
The structural question: does local AI reduce or increase centralization? The obvious answer is reduce — computation moves to the edge, no API dependency, no data leaves the device. The less obvious answer: the model still comes from Google. The weights were trained on Google's data, optimized on Google's hardware, and released on Google's timeline. Local execution doesn't mean local origin. The deployment is distributed. The knowledge production remains concentrated. The threshold for running AI locally just dropped. The threshold for training it didn't move.
