kutaelee/Qwen3.8-27B-RTX5090-128K-Recipe

Qwen3.8-27B on RTX 5090 — 128K Dual-Runtime Agent Recipe

No model weights are hosted here.

This static showcase summarizes a reproducible single-GPU deployment recipe:

• —long / complex / autonomous work → NVFP4 + SGLang;
• —quick / bounded implementation → Q5KM + llama.cpp + MTP3;
• —one resident runtime at a time.

Canonical source and scripts: <https://github.com/kutaelee/qwen38-5090-128k-runtime-recipe>

Reproducibility discussion and issues: <https://github.com/kutaelee/qwen38-5090-128k-runtime-recipe/issues>

The 151.72 tok/s Q5 result did not beat the approximately 69 tok/s SGLang route in one counted autonomous coding trajectory. That is an N=1 trajectory result, not a general ranking of model formats.