At least for qwen3.5, it looks like unsloth has updated their quantization algorithms to avoid bf16. See the march 5th update: https://huggingface.co/unsloth/Q

At least for qwen3.5, it looks like unsloth has updated their quantization algorithms to avoid bf16. See the march 5th update: https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/discussions/31#69a9a3d41fa6f0d3334ff8d9 https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/discussi... I assume they're applying the same technique going forward, but I have no idea how to determine if this is the case.