The Instruct version is specifically tuned to follow complex instructions and features reasoning capabilities (Chain-of-Thought), making it much better at spatial relationships and logical consistency in images.
Yes, because it can follow precise descriptions of equipment, race features, and poses more accurately than the base model.
Due to its massive 80B+ parameter MoE architecture, it requires significant compute resources to run at high speed.