A LocalLLaMA community user says a tool they built to tune copy operations raised prefill speed from 345 to 466 tokens per second in a CPU-offload benchmark.
The user reported the results using llama-bench with CPU offload enabled. The benchmark command used the Qwen3.6-35B-A3B Q8_0 model, a 2,048-token prompt and three repetitions; the user reported prefill speeds of 345 tokens per second before tuning and 466 afterward.
The user said they made a tool to tune copy operations for CPU offloading. The figures come from the user's community post; the article provides no independent verification or further benchmark results.