Google Ships Gemma 4 QAT Checkpoints — Smallest Model Drops From 11.4GB to ~1.1GB

Google / MarkTechPost·high signal

Google DeepMind released Quantization-Aware Training checkpoints for all Gemma 4 sizes, shrinking the smallest model from 11.4GB to 1.1GB (0.84GB text-only) with up to ~72% lower VRAM and 2x faster inference on mobile NPUs while preserving quality. The release includes Q4_0 GGUF for llama.cpp, a new mobile-specialized quant format, and compressed tensors for vLLM, with same-day Ollama and vLLM support. Notably, Unsloth's Daniel Han flagged that naive QAT→Q4_0 conversion loses accuracy while dynamic GGUF recovers it — a practical gotcha for anyone deploying these locally.