Two Free AI Models Build a Virtual V8 Engine—One Wins Big

A tech enthusiast ran the same one-shot prompt through two top locally runnable MoE models to see how they compared, leaving both unattended. DeepSeek V4.1 Flash 1017 (a UD IQ1_M quant) used a single agent for about 60 minutes, generating 250k tokens at ~94 tok/s with a 346k context that never compacted.

Its V8 engine ran on the first try with every button working and looked great — though the cutaway didn’t show much of the inside and it skipped some smaller parts. Qwen 3.8 FN (NVFP4) used a main agent plus five subagents, took around 70 minutes, and produced 443k+ tokens from the main agent and 804k total context with two compactions; two subagents said they’d start writing the file and never did, so the main agent gave up on them.

It built more parts, a better cutaway and even engine sound, but had floating parts, took longer and, in the author’s view, doesn’t look as good. The author stresses this was a one-shot experiment, not the best test, and says they wouldn’t leave Qwen unattended with important work, while DeepSeek seemed better at figuring things out by itself. It ran on Basalt with Quecto Harness on an RTX 5090 + RTX 5060 Ti, Ryzen 9 9950X and 128 GB of DDR5.