On Windows, new LLM inference engines all lose to plain llama.cpp in this RTX 5090 test

No-Bite2555 · reddit · 2026-11-17

A Reddit user with an RTX 5090 (24GB) Windows 11 laptop benchmarked many new inference engines — nInfer, basalt, rusting strata, exLlamaV3 and llama.cpp forks — and found all noticeably slower than his stack of Strata for Qwen 3.8 Flash Next and llama.cpp for everything else. Even using Claude and Codex to tune configs didn't help, leading him to suspect these engines only hit their paper performance on Linux.

Original post →

More from Infra

Infra channel →