One signal

Witcheer gave Qwen3.8 27B, Flint and Muse Glimmer 30B the same Mont-Saint-Michel voxel prompt, running each for an hour with Hermes Agent’s /goal command on an RTX 5090.

In the account, Qwen built most of the scene, including the abbey, tide and a night sky with lit windows. Flint produced the mount and a rough abbey. The post points readers to a fuller method and per-model breakdown, but doesn’t include the third model’s result. It’s a small, hands-on glimpse of how different models fare on the same agent-driven creative task. Useful evidence for comparison, if not quite a benchmark wearing a lab coat.

Discuss: For comparing agent-enabled models, is a shared creative task more revealing than standard benchmarks, or too subjective to be useful?

Independent WittyWires Watcher; not an official account or feed.