Qwen 3.6 27B VLLM Inference Benchmarks: BF16/FP8/NVFP4 Compared
live4evrr · reddit · 2026-07-05
The author tested the VLLM inference performance of three quantization formats (BF16, FP8, NVFP4) for Qwen 3.6 27B on an RTX 6000 Pro Blackwell 96GB GPU. NVFP4 is the fastest but suffers from looping output issues in agent mode. BF16 offers the highest response quality, while FP8 provides the best overall cost-performance ratio. The tests also compared VLLM with llama.cpp: VLLM is faster due to its paged attention mechanism and more stable than llama.cpp (which suffers from random errors requiring prompt resets). The test system used Ubuntu 26.04, vLLM 0.24.0, and CUDA 13.2.
More from Infra
- Nearly half of Claude revenue now flows through cloud marketplace commitments, up from 11% — BenBajarin · 2026-10-26
- Anthropic ARR goes vertical after price cuts: tokens up 4x, $100B+ ARR back in play — GavinSBaker · 2026-10-26
- Alex Banks argues power caps GPU supply, real risk is demand cooling first — thealexbanks · 2026-10-26
- Analyst: Ant has the most fungible AI workload of any model provider today — BenBajarin · 2026-10-26
- Anthropic S-1: training infra edge cut per-task cost 40% and doubled revenue growth — BenBajarin · 2026-10-26
- Rewriting scaling law study: optimal synthetic data share rises from 10% to 32% with compute — heghbalz · 2026-10-26