#2612 Community benchmark: RTX 5090 D, Qwen3.8-Flash-Next IQ3_S at 500K (yarn2)

本文

First RTX 5090 D community report.

- Machine: RTX 5090 D 32 GB, Ryzen 9 9950X, 96 GB DDR5, Windows 11, official Windows release 0.1.44.1 (296b59d2)
- Config: b2 slots, 500K context (yarn2), --kv int8, --resident-experts, --vram-reserve-mib 1500 + elastic, adapt-async + STRATA_PIN_BY_NEED, calibrated expert tier (1/160/0.97)
- Numbers: chat decode 140.2 / agent 162.3 tok/s medians; prefill 4K/16K/32K 4022/5668/5650 tok/s; 3-client burst 106.5 total
- Correctness covered separately: 336K needle retrieval PASS + greedy-output fidelity checks (speed suite itself is speed-only, stated in the entry)
- Reproducible tuning finding included: vram-reserve 12288 -> 1500 on a 32 GB card grew the auto expert cache ~3x, +30% same-prompt decode
- Sanitized: no local paths or usernames (grep -F verified)