Hardware Corner: RTX 3080 10GB, llama.cpp CUDA, Qwen3 8B Q4_K_XL GGUF, decode 74.24 t/s @16k ctx
74.24 tok/s per-request decode (batch 1, ~16384-token context)
Fields
id
benchmarks/hwcorner-rtx-3080-10gb-qwen3-8b-q4kxl-decode-16kMetric
decode_tok_s
Value
74.24
Unit
tok/s per-request decode (batch 1, ~16384-token context)
Model
Qwen3 8B Q4_K XL GGUF
Methodology
THIRD-PARTY MEASURED (hardware-corner.net Definitive GPU Ranking for LLMs, accessed 2026-10-15). llama.cpp llama-bench CUDA 12.8, Ubuntu 24.04, Q4_K_XL GGUF quants, decode tok/s per GPU per model at ~16K context. Per-request batch-1, NOT aggregate. Value is the 16K-context column.
Measured by
third_party
Reproducible
yes
notes
Single-card row at 16K context; same table publishes 32K/65K/131K context columns.
Notes
Single-card row at 16K context; same table publishes 32K/65K/131K context columns.