Gemma 4 · Quantization-Aware Training

QAT Visualizer

Watch precision drop from bf16 to int4 — the model shrinks, the lights stay on. Same quality, half the disk, fully local on your machine.

Checking Gemma 4…
Weight grid · gemma4:e2b-it-qat params: ~2B
precision compression
bf16
16-bit weights
16
bf16
8
int8
6
int6
4
int4 · QAT
0%
Loop
bf16 — fat, glowing, expensive
int4 — small, crisp, on-device
Gemma 4 isn't running yet — this visualizer still works (it's pure animation). Update Ollama, then run: ollama pull gemma4:e2b-it-qat