[Official FAQ] Corrections and follow-up questions welcome.

Full reasoning models can be huge. Distilled variants aim to keep much of the step-by-step behavior in smaller student models that fit consumer hardware better.

Guidelines:

• Tight VRAM: smaller distill + mid quants (e.g. Q4/Q5-class—verify yourself).
• Harder math/code: larger distill or higher precision if you can afford memory and latency.
• Expect longer outputs with visible reasoning traces; your UI must tolerate them.

Leaderboards are hints—run your own tasks. Post hardware + distill name + tokens/s if you can.