[1.0.0]
The initial release, and the one that set ENTITY apart from a generic offline chat app. Rather than run a
desktop-style build on a phone, this version was tuned specifically to the CMF Phone 1's Dimensity 7300.
Inference is pinned to the four Cortex-A78 big cores so the memory-bandwidth-bound decode stays off the
slow A55 efficiency cluster, and the native library is compiled as a single armv8.2-a + dotprod +
KleidiAI backend — arm64 only — instead of the seven generic CPU variants a stock build ships, which
makes the app smaller, quicker to launch, and lighter on RAM. An adaptive-context scheme sizes the KV
window from the model and available memory, which is what lets a 3B model load and generate within roughly
2 GB of free RAM. On top of that runtime sits a proper chat UI with an in-app model picker, smooth token
streaming, a live watts readout, and light/dark/system themes — plus fixes for the early-build problems
that made larger models fail to load and made the model reply with robotic sound effects.
Key components
- Big-core affinity — inference pinned to the Cortex-A78 performance cluster via
sched_setaffinity, with cores chosen by live max frequency (keeps work off the slow A55 cores). - Device-tuned CPU backend — a single
armv8.2-a + dotprod + KleidiAIbuild instead of the seven generic CPU variants; arm64-v8a only (x86 removed) for a smaller, faster, lighter app. - Adaptive context — sizes the context window from model size and free RAM, so a 3B model fits and runs within ~2 GB free (4096 KV).
- Professional chat UI — in-app model picker, smooth token streaming, toggleable stats (tokens, tok/s, TTFT, temperature, power draw in watts, free memory), and light/dark/system themes.
- Black-background white-"E" app icon.
- Fixes over early builds — 3B "error reading file" (KV OOM + a bad on-device copy), model switching, and robotic "beep*boop" replies (via a strict system prompt).
Log in or sign up for Devpost to join the conversation.