Google released Gemma 4 12B, a model designed to fit inside the 16GB RAM threshold most modern consumer laptops already hit.
The model is the latest in Google's Gemma open-weights family, and the headline spec is the 16GB floor — the entry point for most modern laptops, Apple Silicon Macs included. Google says it clears that bar through a new encoding scheme and a token prediction approach that lets the model outperform what 12 billion parameters would ordinarily allow. The pitch is local inference without cloud API calls, a GPU workstation, or a per-query bill.
That pitch is worth taking seriously because the economics of AI inference are genuinely shifting. API costs compound fast at scale, and teams handling sensitive data have real reasons to keep queries off third-party servers. A model that runs well on a developer's laptop — not just technically, but at a usable speed — changes the build-vs.-buy calculus on both fronts. The encoding and token prediction tricks Google is citing are exactly what separates "designed for 16GB" from "technically boots on 16GB."
Meta's Llama family and Mistral's models have been making the same "runs locally, punches above its weight" claim for well over a year. Google arrives with distribution advantages — Gemma fits naturally into its tooling and Android-adjacent workflows — but the benchmark that matters isn't the press release; it's whether developers swap their API calls for a local model file.