ParentFull threadkarolist·35B models need about 22G vram with Q6 quantization, you should be able to run 70B models with low QView on HN