I've been experimenting with local models on an M5 Max MBP w/ 128GB of RAM since March of this year. Generally I've had very good results. Where things were lacking initially was with tool calling and the need to rely on tool calling for functionality like web search, which is otherwise well integrated in the cloud models. There is also a lot more work required on the harness side, however at this point (August 2026) there is not only much better tool calling in local models, but community supported projects have built good harnesses. Pi.dev and OpenCode + a SOTA local model is /very/ /very/ capable these days. Using LM Studio's built-in chat with a decent system prompt and proper tuning with local models is /very/ capable these days. Cloud models are still better, but it's the "harness" (not in the desktop app, but in the backend) that makes it so for the most part.