So far, I haven't been happy with any of the smaller coding models, they just don't compare to claude/codex.