Hardware Corner Qwen3.8-27B Q4_K_S: RTX 3090 24GB tg 33.95 t/s @64k ctx
33.95 tok/s per-request decode (batch 1, 64k context)
Fields
id
benchmarks/hc38-rtx-3090-qwen38-27b-tg-64kMetric
decode_tok_s
Value
33.95
Unit
tok/s per-request decode (batch 1, 64k context)
Model
Qwen3.8 27B Q4_K Small GGUF
Methodology
THIRD-PARTY MEASURED (hardware-corner.net 'We Tested Qwen3.8 27B', accessed 2026-10-15). llama.cpp build 153d324bc (10364), Ubuntu 24.04, NVIDIA driver 595.58.03, CUDA 12.8, Qwen3.8-27B Q4_K Small GGUF (16.68 GiB, 27.32B params; llama.cpp reports arch 'qwen35'), Flash Attention on, MTP off, full GPU offload. RTX 3090 24GB. pp = prompt-processing tok/s, tg = token-generation tok/s at stated prefilled context. Per-request batch-1.
Measured by
third_party
Reproducible
yes
notes
64k is the practical context ceiling on 24 GB (measured 22 GB VRAM).
Notes
64k is the practical context ceiling on 24 GB (measured 22 GB VRAM).