On an M5 Max with the network off, calling fm from the terminal gave about 46 tok/s, with the first token in 0.43 s. The context limit is about 8K (8,026 tokens worked, 8,095 errored), not the 4K people quote. In a Chinese typo test it missed 3 of 12. Post ↗
Counted on an in-store unit with the keyboard up: the iPhone Air shows 7 messages, the Duo cover screen only 4. I also made a web page that simulates it on the phone you already have. Post ↗
At inference time, nudge L14·N8949 up and L13·N7956 down. At 0.5–0.75× it turns “evil” but stays fluent; at 1× it starts repeating itself; at 2× it’s gibberish. I reproduced it once and it held. Post ↗