Unsloth Qwen3.6 GGUFs Lag CPU Quants

Hot take: this looks less like a one-off glitch and more like a quantization or runtime-tuning tradeoff that becomes obvious on CPU-only inference.

• –The reported gap is consistent across both IQ4_NL and IQ4_XS variants, which points to a systematic difference rather than a single bad file.
• –The user’s environment is CPU-only llama.cpp, so the result may not translate to GPU-backed or different-runtime deployments.
• –Unsloth’s own docs emphasize benchmarked Dynamic GGUFs and note that some accuracy-oriented choices can cost inference speed, so this could be an intended tradeoff rather than a bug.
• –The first-followup latency is also worse, which suggests the issue may involve prompt processing or cache behavior, not just raw decode throughput.
• –If reproducible, the next thing to compare is the exact quant recipe, llama.cpp build flags, context settings, and chat template behavior. This is an inference from the report, not something the post proves directly.