verifiedmedium · 0.70updated 2026-10-14

RTX 5090 llama.cpp: Qwen3 30B-A3B (MoE) Q4_K_XL decode at 4k ctx (Hardware Corner)

234.3 tok/s (tg128, batch 1, per-request)

Fields

benchmarks/hc-rtx5090-qwen3-30b-a3b-q4kxl-4k-tg128

Metric

decode_tok_s

Value

234.3

Unit

tok/s (tg128, batch 1, per-request)

Model

Qwen3 30B-A3B (MoE)

Methodology

Hardware Corner (Allan Witt) llama-bench build 466c1911 on RTX 5090 32GB GDDR7 (EPYC 7642 host, Ubuntu 24.04). Qwen3 30B-A3B (MoE) Q4_K_XL, -fa 1 -ngl 99, -p 1024/2048 -n 128 at context depth 4k. Per-request batch-1, NOT aggregate.

Measured by

third_party

Reproducible

yes

notes

PP at same ctx: 6630.42 t/s. Full ctx sweep 4k..131k in source.

Notes

PP at same ctx: 6630.42 t/s. Full ctx sweep 4k..131k in source.

Cross-references

RTX 5090 llama.cpp: Qwen3 30B-A3B (MoE) Q4_K_XL decode at 4k ctx (Hardware Corner)

Fields

Notes

Cross-references

Accelerators (1)

Engine (1)

Format (1)

Interconnect (1)