Show HN: Maple-Preview – ternary 20B MoE running at 120 tok/s on a iPhone

A tiny internet stampede broke out after DeepGrove showed off Maple-Preview, an open-source AI model it says can run shockingly fast on everyday devices — including an iPhone — while still tackling seriously hard math. On paper, it sounds like sci-fi: a model small enough to fit in about 5.3 GB, fast enough to spit out text at 127 tokens per second on an iPhone and more than 200 on a Mac mini, and flashy enough to solve an International Mathematical Olympiad problem. Unsurprisingly, some readers were instantly sold. One early reaction was the pure, wholesome “This is cool!”, while another cheered, “Edge is edging closer!” — basically the local-AI fan club declaring the future has entered the chat.

But this being the internet, the applause lasted about three seconds before the side-eye began. The biggest complaint? Speed is nice, but what if it’s fast and wrong? One commenter said the model “hallucinates knowledge quite aggressively,” accusing the online demo of hiding the problem behind web search tools. Another was even harsher, begging so-called “small” AI tools to stop being “confidently very incorrect.” And then came the bonus drama: one user noticed several low-karma accounts praising the launch and flatly called it “Suspicious.” So now the real story isn’t just that your phone may soon run powerful AI offline — it’s that the crowd is split between “holy wow” and “nice demo, but is this thing bluffing at light speed?”