arm64-v8a · Android 13+ · Apache-2.0

Two apps, one runtime

ENTITY is the assistant. ENTITY Bench is the same three-arm ablation as a standalone tool, so any arm64 phone can reproduce or challenge the findings - and contribute a row to the public dataset.

ENTITY v3.7.1

on-device LLM chat

Download APK · 10.0 MB

file: ENTITY-v26-review-fixes-20260814-release.apk
sha-256: e97ec05fa29f1184db0f5c5fb3430e8db091fa4a80b6d7a62c2f3dc41d1243d8
requires: arm64-v8a, Android 13+

Upgrading from v3.6.2 or earlier? every build from v3.7.0 on is signed with a new certificate, so an in-place update fails with INSTALL_FAILED_UPDATE_INCOMPATIBLE and you have to uninstall first - which deletes saved conversations. Export anything you want with SHARE CHAT before you do. A fresh install is unaffected, and so is an update from v3.7.0 - that build carries the same certificate as this one. A certificate mismatch normally means a tampered build, so this is stated here rather than left to surprise you; the reason is entry 11 on the journey page.

ENTITY Bench v2.2.1

on-device LLM benchmark

Download APK · 9.9 MB

file: ENTITY-Bench-v2.2.1-release.apk
sha-256: 5eaf3db5ae6908626ea3e9bef6d08f587ac181d1109edc796516bc8e1deb9a9d
requires: arm64-v8a, Android 13+

Same certificate change as the assistant. Upgrading over v2.1.1 fails and needs an uninstall first, which deletes saved benchmark history - export each result you want to keep to CSV beforehand. A fresh install is unaffected, and so is an update from v2.2.0.

Ninety seconds from clone to chatting

git clone https://github.com/kkjjkamal123/ENTITY---Arm-Create-AI-Optimization-Challenge.git
cd ENTITY---Arm-Create-AI-Optimization-Challenge
adb install -r apk/ENTITY-v26-review-fixes-20260814-release.apk

Then on the phone: pick a model from the in-app catalog or import a GGUF you already have. Prefer Q4_0 or Q8_0 - the only two types that reach Arm's KleidiAI kernels (why); the recommended starting point is Llama-3.2-1B-Instruct-Q4_0.gguf. Leave Auto on, then open BENCHMARK and run the ablation unplugged to see power and tokens per watt.

ENTITY - the assistant

Fully offline chat for runnable GGUF models (Llama 3.2 1B / 3B and others).
Streaming replies with Stop, Markdown and LaTeX rendering, copy and regenerate.
Persistent local conversations - restore, rename, switch, delete.
Auto mode plus manual temperature, top-k, top-p, length, context and threads.
Live stats bar and seven-series graph: tokens, rate, TTFT, temperature, watts, CPU, memory.
In-app three-arm benchmark with history, CSV export and core-placement verdict.
A recommended model for this phone, from a 60 ms probe that needs no model downloaded.
Model cards report the KleidiAI coverage measured from the file's own tensor table.
Long-press any reply for its token accounting: prompt read, tokens written, both rates.

Before you download

The catalog spans seven vendors and 0.36B to 7B. Each entry is judged against memory the phone reports free right now - not installed RAM, which counts capacity already spent on everything else open - and, once the probe has run, carries an estimated generation rate and time to first token. The reasoning is on the optimizations page.

What the file actually contains

A GGUF named Q4_0 is not uniformly Q4_0. ENTITY parses the tensor table and reports the fraction of weights that reach Arm's kernels, naming the largest tensor that does not - on the reference Llama-3.2-1B-Q4_0 that is 76%, and the missing quarter includes the output projection.

What that answer cost

Prompt and generated tokens are reported apart, never as one rate. They are produced by different regimes - the prompt goes through in batches and is compute-bound, each generated token re-reads every weight and is bandwidth-bound. Averaging them is why speed claims about local models are so often incomparable. Timings cover model execution only, so a slow redraw cannot pass as a slow model.

ENTITY chat - 1 / 9

Live metrics while it generates

100 tok - 16.4 t/s - TTFT 299 ms - 33.0C - 4.34 W - CPU 47% - 1.2 GB, plus the selectable colour graph, the one colour surface in the app.

ENTITY chat - 2 / 9

One drawer, everything

v3.6.2, conversation history, model selector, BENCHMARK / SHARE CHAT / BENCHMARK HISTORY, SETTINGS and ABOUT.

ENTITY chat - 3 / 9

A catalog sized to your phone

On-device files plus a catalog ranked for this SoC, with KLEIDIAI / NO KLEIDIAI and GOOD FIT / FITS chips per row.

ENTITY chat - 4 / 9

The full model readout

KLEIDIAI ACTIVE - Q4_0, 773 MB, trained context 131,072, running context 4096 auto-fit to free RAM. Compute: CPU, 4 perf cores, dotprod, KleidiAI active.

ENTITY chat - 5 / 9

Benchmark from inside chat

Runs per config, sustained thermal test, PP 512 / TG 128, and a cooldown before every pass.

ENTITY chat - 6 / 9

The runtime, in its own words

BIG-CORE AFFINITY, RUNTIME-SELECTED CPU BACKEND, ADAPTIVE CONTEXT, AUTO MODE, THERMAL-AWARE GUARD, KV SESSION CACHE.

ENTITY chat - 7 / 9

Yours to theme

Theme (System / Light / Dark), Monochrome / Colour palette, text size, and which live-metric series the graph draws.

ENTITY chat - 8 / 9

Auto, or exactly your way

Auto (optimized) plus manual sliders, Efficiency mode, and CORE PLACEMENT: AUTO / PERF CORES / SCHEDULER - pinning is a setting, not an assumption.

ENTITY chat - 9 / 9

Every run, kept on-device

Benchmark history saved locally: +85% decode vs naive, 1 run, 3-arm, unplugged.

ENTITY Bench - the experiment, standalone

No chat. Import a model, run the ablation, get an attributed result: naive, threads-only, auto - plus the optional efficiency-core arm, SUSTAINED mode (who throttles first once the SoC is hot) and SWEEP mode (every thread width, pinned and scheduler-placed).