Hardware Corner — We Tested Qwen3.8 27B (llama.cpp b10364, CUDA 12.8: RTX 3090/4090/5090, dual RTX 5060 Ti, M5 Max 128GB, NVIDIA GB10 pp/tg vs context)
benchmark
Fields
id
sources/hardware-corner-qwen38-27b-hardware-testsPublisher
Hardware Corner
Kind
benchmark
Accessed
2026-10-15
notes
In-lab llama.cpp build 153d324bc (10364), Ubuntu 24.04, driver 595.58.03, CUDA 12.8 (GB10: CUDA 13.2, driver 595.71.05), Qwen3.8-27B Q4_K Small GGUF 16.68 GiB, FA on, MTP off, full GPU offload. pp and tg tok/s vs context 4k-256k. Batch-1 per-request. Third-party measured.
Notes
In-lab llama.cpp build 153d324bc (10364), Ubuntu 24.04, driver 595.58.03, CUDA 12.8 (GB10: CUDA 13.2, driver 595.71.05), Qwen3.8-27B Q4_K Small GGUF 16.68 GiB, FA on, MTP off, full GPU offload. pp and tg tok/s vs context 4k-256k. Batch-1 per-request. Third-party measured.
Cross-references
This record references no other records.