Model capability is what harnesses expose, I have experimented with pi using local mlx models - several 27/35B quantisations (Gemma, Qwen, Flint) and the reality is that they are much weaker than Sonnet, let alone Opus. Of course they are "free" (if you have a $3000 mac) but they are slow and dumb, useful for very routine things but comparatively bad at problem solving and trouble shooting.