Qwen3.8-Flash-Next on 2x3090: Decode Boosted From 25 to 41 t/s With Cache + MTP
Extension-Bid-639 · reddit · 2026-09-04
A Reddit user details pushing Qwen3.8-Flash-Next (261k context, all 48 expert layers in host RAM) from 17 to 37-41 t/s decode on 2xRTX 3090 + 188GB DDR4: switching to UD-Q4KXL for more expert-cache slots (90-92% hit rate), stacking MTP on top (50-58% acceptance, 49 t/s on code emission), fixing loader page faults (13→2 min load), patching an out-of-bounds bug in the cache PR, and discovering DDR4 thermal throttling at 78°C.
Related event: Tuning Qwen3.8-Flash-Next on 2x3090: From 17 to 41 t/s Decoding(2 posts)→
More from Infra
- Google open-sources TorchTPU, joining Amazon in loosening Nvidia's grip on PyTorch developers — gaganghotra_ · 2026-10-21
- FastVideo open-sources MiniMax FastH3 Omni with training code, 12x faster on Blackwell — multimodalart · 2026-10-21
- Anthropic's draft IPO filing: $16.3B H1 revenue, $80B net loss, $750B compute plans — kimmonismus · 2026-10-21
- dphnAI's distributed inference network served 32B DeepSeek tokens in 24h, ~6% of Venice's daily volume — Ronangmi · 2026-10-21
- Jensen Huang: every company will have two factories, one real and one digital — round · 2026-10-21
- DeepSeek V4 called 'god-tier infra, 4th-tier ML': 7 tricks stacked, 20 patches to fix them — teortaxesTex · 2026-10-21