GitHub Repository

Android 16 fork. AI as a platform primitive. Twelve capabilities, one shared runtime, every app. OEM-pluggable. Apache 2.0.

10 starsShell

JibarOS, a shared inference runtime for Android

Name: JibarOS, a shared inference runtime for Android
Availability: InStock
Author: rafaelvalle03

by rafaelvalle03·Apr 23, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●●BangerZero to OneBig BrainBold Bet

Shared inference runtime at the OS level saves RAM compared to per-app model bundling.

Strengths

•System-wide Binder AIDL exposes 12 capabilities, avoiding per-app model bundles.
•Centralized model residency prevents duplicate RAM usage across concurrent applications.
•Native daemon handles scheduling and fairness without requiring app-side changes.

Weaknesses

•Requires flashing a custom Android fork, limiting immediate adoption and testing.
•Only 10 commits; needs more backend integrations beyond current proofs.

Post Description

JibarOS is an Android 16 fork I’ve been building to explore a shared runtime for on-device inference.

It adds: a system service in system_server a native daemon pluggable inference backends a Binder AIDL for capability-based calls across text, audio, and vision The goal is to centralize things that are otherwise handled independently by apps, like model residency, scheduling, fairness, and backend routing.

The current interface exposes 12 capabilities including completion, translation, rerank, embeddings, transcription, synthesis, VAD, OCR, detection, and description.

Repo: https://github.com/Jibar-OS/JibarOS

Interested in feedback from anyone who has worked on Android framework/services, ML runtimes, or device-level resource scheduling.