That’s mostly because current open models are way too heavy and too general for coding. If the model is compact enough to run well on normal Apple Silicon, local stops being an overnight-only toy and you also lose the third-party dependency. That’s the bet behind rig.ai.