Gemma 4 QAT models quantized to NVFP4 with GPTQ+iMatrix and FP8-calibrated KV cache using multilingual and tool-use data.