A EUR 1,600 local rig for running Qwen3.8-Flash-Next at 70-90 tokens/sec

On 2026-10-24 02:22, a comparison post appeared in the Reddit LocalLLaMA community. The author describes himself as a developer and, in the same breath, as the developer of Basalt, one of the engines being measured. He writes that after roughly five excruciating hours of setup he benchmarked all of the Qwen 3.8 FN engines he had been hearing people talk about, and published the comparison data. To keep things as fair as possible, he used the same checkpoint, GPU and requests across every run, and configured each engine the way its own documentation recommends. The hardware was a single 5090 paired with 128GB of memory. He also notes that the post had some AI cleanup applied to his own wording, and that the details of how the tests were run are at the bottom of the post. The existing material does not say how many engines were included, which engines besides Basalt were tested, or what the measured scores were, so the current picture stops at the publication of the comparison itself, with no follow-up coverage reported yet.