I ran all the popular Qwen 3.8 FN engines on a single 5090 + 128GB of RAM, here are the numbers

Before you read: I'm the developer of Basalt, one of the engines benchmarked below. I tried to keep this as fair as possible: same checkpoint, GPU and requests, and every engine set up the way its own docs recommend. How it was run is at the bottom. This post has had some AI cleanup from my o

中文全文翻译正在补齐，可先查看下方英文原文。

Before you read: I'm the developer of Basalt, one of the engines benchmarked below. I tried to keep this as fair as possible: same checkpoint, GPU and requests, and every engine set up the way its own docs recommend. How it was run is at the bottom. This post has had some AI cleanup from my o