Seinari on iPhone: a ternary 2B model chatting entirely on the phone
Three questions, three answers, in real time. Each reply appears in well under a second because the phone generates about 200–235 word pieces (tokens) per second. The number at the top of the screen is the speed of the last reply, measured by the app.
In this chat: 203, 235 and 197 tokens/s, with the first word after 103–137 ms. No internet connection or server is involved in answering; the model (PrismML's Ternary Bonsai 2 2B, 429 MB) runs on the phone's GPU in Seinari's own Metal inference engine. Its output is identical, token for token, to plain decoding; the speed comes from exact speculative decoding.
For scale: the model's maker publishes about 120 tokens/s for this model on iPhone, and plain decoding on this phone runs at about 163.
About the recording: iPhone 18 Pro Max (A20 Pro) in a cloud device lab, 2026-10-21. The questions are typed into the message bar by the app itself (a scripted demo), not by hand. This is a 30-second stretch of a longer real-time recording, unedited and not sped up; only the app launch and model download before it are trimmed.
Raw benchmark: 268 tokens/s
The same engine on the same phone, running our eight-prompt speed benchmark, averages 268 tokens/s (individual prompts 215–428). Uncut 87-second benchmark recording.