Gemma 4 QAT Q4_0 Bench on Strix Halo These are Google's official Gemma 4 QAT Q4_0 GGUF models, served locally through llama.cpp Vulkan/RADV on a Strix Halo APU. QAT means quantization-aware training . Instead of taking a normal model and quantizing it only after training, the model is trained or adapted while accounting for the lower-precision format it will run in. The goal is to make a small Q4 model keep more of the original model's behavior than a simple post-training quantization. Ho...