I tested Qwen3.8 27B on my DGX Spark in the BF16, Q6, and NVFP4 versions. Here are the results: **Overall** • NVFP4 / vLLM: 🥇 Winner • Q6_K / llama.cpp: 🥈 Fast-start option • BF16 / vLLM: 🥉 Prefill specialist Artifact size • NVFP4 / vLLM: 23.45 GB • Q6_K / llama.cpp: 22.88 GB • BF16 / vLLM: 55.59 GB Model/resident memory • NVFP4 / vLLM: 21.34 GiB • Q6_K / llama.cpp: ~36 GiB • BF16 / vLLM: 51.1 GiB Cold API readiness • NVFP4 / vLLM: 386 s • Q6_K / llama.cpp: 6.23 s • BF16 / vLLM: ~473 s **GSM8K** • NVFP4 / vLLM: 93% • Q6_K / llama.cpp: 95% • BF16 / vLLM: 93% **IFEval** • NVFP4 / vLLM: 83% • Q6_K / llama.cpp: 80% • BF16 / vLLM: — Mean TTFT • NVFP4 / vLLM: 193 ms • Q6_K / llama.cpp: 455 ms • BF16 / vLLM: 485 ms Median latency • NVFP4 / vLLM: 1.80 s • Q6_K / llama.cpp: 2.48 s • BF16 / vLLM: 4.53 s Single-stream decode • NVFP4 / vLLM: 10.86 tok/s • Q6_K / llama.cpp: 8.18 tok/s • BF16 / vLLM: 4.30 tok/s C2 aggregate • NVFP4 / vLLM: 21.29 tok/s • Q6_K / llama.cpp: 15.43 tok/s • BF16 / vLLM: 8.58 tok/s C4 aggregate • NVFP4 / vLLM: 41.17 tok/s • Q6_K / llama.cpp: 11.17 tok/s • BF16 / vLLM: 16.93 tok/s C6 aggregate • NVFP4 / vLLM: 60.00 tok/s • Q6_K / llama.cpp: 32.39 tok/s • BF16 / vLLM: 24.89 tok/s 8K prefill proxy • NVFP4 / vLLM: 1,308 tok/s • Q6_K / llama.cpp: 648 tok/s • BF16 / vLLM: 10,228 tok/s Best use • NVFP4 / vLLM: Default deployment • Q6_K / llama.cpp: Instant startup • BF16 / vLLM: Prefill-heavy workloads Bottom line: NVFP4 offers the best combination of capability, memory efficiency, response latency, and multi-user throughput.