https://huggingface.co/mradermacher/gemma-4-31B-it-qat-q4_0-unquantized-i1-GGUF/tree/main https://huggingface.co/mradermacher/gemma-4-31B-it-qat-q4_0-unquantized-GGUF/tree/main I waited a bit before asking this. I have 3060 12GB and 32GB ddr3 RAM. I'm currently using an old version of unsloth's gemma-4-31B-it-UD-IQ3_XXS.gguf which is 11.8GB. With override ffn_down tensors, I can run 16k bf16 context at about 1.3 tk/s last time I used it. When I use the bf16 mmproj I offload it to CPU. Ove...