Phone-Speed Leaderboard
Real on-device LLM benchmarks, submitted from the TokForge app on real phones, now including iPhone. Ranked by decode speed in tokens per second. No simulators and no cloud, just what these devices actually do.
How to read this board
Decode t/s
How fast the phone writes its reply, in tokens per second. A token is roughly three quarters of a word. 15 or more feels like comfortable reading speed, 8 to 15 is usable, under 8 feels slow.
Prefill t/s
How fast the phone reads your prompt before it starts answering. Higher means less waiting before the first word appears. A dash means that run did not measure prefill.
TTFT
Time to first token: the pause between sending a message and seeing the first word of the answer, in milliseconds. Reported by newer app versions, so a dash just means an older submission.
Size classes
Models are grouped by total parameter count: Under 2B, 2-4B (2B up to 5B), 7-9B (5B up to 10B), and 14B+ (10B and up). MoE (mixture-of-experts) models get their own tab because only a fraction of their parameters is active per token, so they behave differently from dense models of the same total size. Models we cannot size yet appear under Other.
What counts as a run
Every score comes from the TokForge app's built-in benchmark on a real phone: a fixed prompt, a fixed reply length, and several timed passes. Each row shows the best submitted score for that phone chip, model, backend, and quantization. Scores are signed by the app before they are sent, which keeps drive-by fake numbers off the board and carries no personal data.
Submissions come straight from the app's Forge/Benchmark screen (opt-in) and are anti-spam signed. Raw API: /v1/board · standalone dashboard: leaderboard.tokforge.ai. Decode tokens/sec is sustained generation speed; higher is faster.