We also noticed a surprising capability in non-coding reasoning, despite the model being trained entirely on code. We benchmarked against models trained for reasoning tasks and replit-code-v1-3b performed incredibly well.

→ View original post on X — @replit,