RTX 5070 Ti llama.cpp: gpt-oss 20B (MXFP4) decode at 64k (Hardware Corner per-card page)
91.3 tok/s (token generation, per-request, batch 1)
Fields
id
benchmarks/hc26-5070ti-gptoss-20b-tg-64kMetric
decode_tok_s
Value
91.3
Unit
tok/s (token generation, per-request, batch 1)
Model
gpt-oss 20B (MXFP4)
Methodology
Hardware Corner per-card LLM benchmark page (March 2026 sweep). Same lab method as their RTX 5090 llama-bench series already ingested: llama.cpp llama-bench, Ubuntu 24.04, CUDA 12.8, Q4_K GGUF (gpt-oss 20B labeled MXFP4 on the page; filed under gguf-mxfp4-moe, the nearest schema format). Context depth 64k. Per-request, batch size 1, NOT aggregate across concurrent requests. The page does not print the llama.cpp commit or the exact prompt/output lengths; the sibling 5090 series used build 466c1911 with -fa 1 -ngl 99. That protocol is the likely one and is NOT asserted as measured on this card. Token generation 91.3 tok/s.
Measured by
third_party
Reproducible
yes
notes
Transcribed from the Hardware Corner RTX 5070 Ti per-card page (header: Updated Benchmarks March 2026), not from the Dec 2025 ranking table. The two pages disagree on the same cell: Qwen3 8B decode at 16k is 87.5 here and 127.72 on the ranking, so this record is status: contested. 16 GB GDDR7, 896 GB/s, 300 W. Page states gpt-oss 20B MXFP4 runs at a full 128k context in VRAM and that 30B+ models do not fit.
Notes
Transcribed from the Hardware Corner RTX 5070 Ti per-card page (header: Updated Benchmarks March 2026), not from the Dec 2025 ranking table. The two pages disagree on the same cell: Qwen3 8B decode at 16k is 87.5 here and 127.72 on the ranking, so this record is status: contested. 16 GB GDDR7, 896 GB/s, 300 W. Page states gpt-oss 20B MXFP4 runs at a full 128k context in VRAM and that 30B+ models do not fit.