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 fast does a phone actually run a local LLM? This page answers that with measurements instead of estimates. Every row is a benchmark run submitted from the TokForge app on a real handset: how many tokens per second the phone writes (decode), how fast it reads the prompt (prefill), how long the wait is before the first word appears (time to first token), and the exact model, quantization and backend that produced the number.

If the question is which phone is fastest for local AI, whether an iPhone keeps up with an Android flagship on Qwen or Llama, what an SM8850, an SM8750, an MT6899 or an Apple A17 Pro does on a small model, or whether a 9B model stays usable on a phone at all, the table below is the evidence for the devices people have actually tested. Speeds vary with the model, the quantization, the backend and how warm the phone is, so every row carries all four rather than a single headline figure.

The board is live. The full dataset is also baked into this page as static HTML with a dated snapshot, so it can be read and quoted without running JavaScript, and it is available as JSON from the public API linked at the foot of the page.

The full benchmark dataset

Every run the public API returns, unfiltered, baked into this page as static HTML. This is the citable copy: it does not need JavaScript, it carries an explicit measurement date per row, and it is refreshed daily from the same source the live board reads. The table scrolls sideways: past the model column each row also carries parameter count, quantization, backend, prefill speed, decode speed, time to first token, RAM tier, OS build, the method the passes were combined with, the app version and the measurement date.

Snapshot: 2026-08-22 (UTC). 64 benchmark runs from 44 devices, 28 chips and 37 model builds. Oldest run measured 2026-06-10, newest 2026-08-22, submitted from TokForge 1.1.1. Data collected with the TokForge app on physical devices, no simulators and no cloud inference. The App version column carries the exact build behind every row. Refreshed daily from the live board.

How these numbers were produced

Each score comes from the benchmark built into the TokForge app, run on a physical phone that the owner chose to submit from. The prompt is fixed, the reply length is capped so every run generates the same amount of work (128 tokens on Android, 96 on iPhone), and the app runs several timed passes rather than one. The Method column records how those passes were combined: best3 is the best of three passes with the prefill figure taken from that same pass, median3 is the median of three, and warmup1_median2 is one warmup pass followed by the median of two. Rows submitted by app versions that predate the tag show a dash. Submissions are signed by the app before they are sent, which keeps invented numbers off the board, and they carry no personal data.

Each row is the best submitted score for one combination of phone chip, model, backend and quantization, so a device can appear more than once with different models. TTFT, peak memory, OS version and the method tag arrived with later app versions, so a dash means the run predates that field rather than a failed measurement. Nothing here is filtered for flattery: slow rows, old builds and odd model names all stay in.

To cite it: TokForge On-Device LLM Benchmark Dataset, snapshot date as shown above, https://tokforge.ai/leaderboard/. Machine-readable JSON for the same rows is at leaderboard.tokforge.ai/v1/board.

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.