Hardware Corner: RTX 3090 Ti 24GB, llama.cpp CUDA, gpt-oss 20B Q4_K XL GGUF, decode 94.91 t/s @65k ctx
94.91 tok/s per-request decode (batch 1, ~65536-token context)
Fields
id
benchmarks/hwcorner-rtx-3090-ti-gptoss-20b-q4kxl-decode-65kMetric
decode_tok_s
Value
94.91
Unit
tok/s per-request decode (batch 1, ~65536-token context)
Model
gpt-oss 20B Q4_K XL GGUF
Methodology
THIRD-PARTY MEASURED (hardware-corner.net Definitive GPU Ranking for LLMs, page updated 2025-12-09, accessed 2026-10-16). llama.cpp llama-bench on Ubuntu 24.04, CUDA 12.8, Q4_K_XL GGUF (nearest repo format gguf-q4-k-m), single GPU, full offload. Metric is per-request decode tok/s (tg) at a filled ~65536-token (65k) context, batch 1, NOT aggregate across concurrent requests. Value is the 65k column of the token-generation table for gpt-oss 20B.
Measured by
third_party
Reproducible
yes
notes
Single-GPU batch-1 row from the Hardware Corner in-lab ranking table. Q4_K_XL is their published quant label; filed under gguf-q4-k-m as the nearest schema format, same convention as the 16k rows already in this repo. Per-request, not an aggregate serving number. The page does not publish the exact llama.cpp commit hash for this table.
Notes
Single-GPU batch-1 row from the Hardware Corner in-lab ranking table. Q4_K_XL is their published quant label; filed under gguf-q4-k-m as the nearest schema format, same convention as the 16k rows already in this repo. Per-request, not an aggregate serving number. The page does not publish the exact llama.cpp commit hash for this table.