Qwen-3.8-Flash-Next on 1x RTX 5090: TG=50 t/s, PP=2300 t/s - with FreeToken

I was somewhat disappointed with the performance of Qwen 3.8 Flash Next on my single RTX 5090 using llama.cpp. One issue is that llama.cpp still has no Expert Caching implemented for MoE models. There are various PRs and discussions (see here, for example), but nothing is merged yet. Then I stumbled

中文全文翻译正在补齐，可先查看下方英文原文。

I was somewhat disappointed with the performance of Qwen 3.8 Flash Next on my single RTX 5090 using llama.cpp. One issue is that llama.cpp still has no Expert Caching implemented for MoE models. There are various PRs and discussions (see here, for example), but nothing is merged yet. Then I stumbled