This Reddit post introduces a speculative decoding technique for large language models that exceed GPU memory capacity. The method utilizes a 2-bit quantized copy in VRAM combined with an exact residual in system RAM to verify predictions. The author reports significant cost savings in their production application and encourages further exploration.