Hardware Corner: RTX 5060 Ti 16GB, llama.cpp CUDA, Qwen3 8B Q4_K GGUF, token generation (tg) 69.2 t/s @4k ctx
69.2 tok/s per-request decode (batch 1, ~4096-token context)
Fields
id
benchmarks/hwcorner-rtx-5060-ti-qwen3-8b-q4k-tg-4kMetric
decode_tok_s
Value
69.2
Unit
tok/s per-request decode (batch 1, ~4096-token context)
Model
Qwen3 8B Q4_K GGUF
Methodology
THIRD-PARTY MEASURED (hardware-corner.net per-GPU LLM page, March 2026 update, accessed 2026-10-16). llama.cpp CUDA, single GPU, batch 1. RTX 5060 Ti 16GB, Qwen3 8B Q4_K GGUF. Grid cell is token generation (tg) at the 4k-context column (~4096 tokens). Per-request, not summed across concurrent requests. Q4_K is the lab label; nearest repo format is gguf-q4-k-m. Prompt length inside the column is the context tier, not a separately published pp512.
Measured by
third_party
Reproducible
yes
notes
Context-sweep companion to the 16k decode rows from the same lab ranking page. Hero callouts that disagree with the grid (4070 Ti SUPER 84.3 t/s copies the PRO 5000 callout) are not used; this is the grid cell.
Notes
Context-sweep companion to the 16k decode rows from the same lab ranking page. Hero callouts that disagree with the grid (4070 Ti SUPER 84.3 t/s copies the PRO 5000 callout) are not used; this is the grid cell.