虚拟机跑Strata + Qwen3.8 Flash-Next（直通RTX PRO 5000 48G 分配128G内存），252K上下文，80~100TOKEN之间，用的是Unsloth UD-IQ4_XS量化，单并发

[Apache 0]

以下是经过测试后的参数，感觉还有空间，不知道如何调了，有同配置的大佬可以多交流
"exe": "/opt/strata/engine/strata",
"args": [
"--pack",
"/data/packs/unsloth-ud-iq4_xs",
"--native",
"/data/models/unsloth-UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf",
"--expert-profile",
"/opt/strata/data/expert-profile.bin",
"--expert-cache", "auto",
"--prefill", "auto",
"--spec", "4",
"--spec-min-p", "0.5",
"--mtp", "/data/mtp/rt",
"--max-context", "262144",
"--kv", "int8",
"--resident-budget-gib", "55",
"--vision",
"--vram-reserve-mib", "700",
"--conversation-cache-mib", "20480",
"--conversation-cache-slots", "6",
"--conversation-cache-min-free-mib", "13824"

使用deepseek harness调用，非常丝滑，用了几天下来，也没有出现任务中断情况，比Qwen3.8 27B FP8更快，更聪明