Qwen3.8 Flash Next hits 262 tok/s on 16GB VRAM

A LocalLLaMA benchmark shows Qwen3.8-Flash-Next IQ3_S running on an RTX 5080 with 16GB VRAM and 55GB RAM, reaching about 5,600 tok/s prompt processing and 262 tok/s generation at 262K context. In an agent coding session, 99.5% of prompt tokens hit cache, with 0.2s median time to first token.

• Developer posts update on local Yandex 80B finetune2026-10-05
• Blogger says self-hosting AI saves no money but sticks with it2026-10-03
• Local AI costs €0.12/hour: is paying for frontier models cheaper?2026-10-01
• Would a $1k Qwen3.8-27B Dedicated Chip Have a Market?2026-09-21
• Community discusses Qwen4 expectations and model size needs2026-10-26