Introduction · Performance · How It Works

The AI that lives inside your phone.

Muse is a fully offline, on-device AI assistant for Android. No cloud. No subscriptions. No data leaving your device — ever. This page walks you through exactly what it does, how fast it is, and why it's fundamentally different from every other AI app.

Data sent to cloud

0ms

Network latency

AI models available

Free

No subscription

What is Muse?

A private AI that runs on your hardware.

Every mainstream AI app — ChatGPT, Gemini, Copilot — sends your messages to a server. Muse doesn't. It runs a full language model on the phone itself, so your words never leave your device.

🔒

Zero Cloud Exposure

Conversations are processed entirely on your CPU/GPU. There is no server to breach because there is no server.

✈️

Works Fully Offline

Airplane mode? No Wi-Fi? Doesn't matter. Muse runs on the device hardware — no internet needed at any point.

🧠

Persistent Memory

Muse remembers facts about you across sessions — preferences, context, goals — stored locally and injected into every prompt.

🎙️

Offline Voice Mode

Full speech-to-text and Piper TTS synthesis, both running on-device. Voice chat with zero audio sent anywhere.

🌐

Mesh Inference

Other phones on the same Wi-Fi are auto-discovered and share inference work. More phones = more compute, no setup needed.

⚡

Action Engine

Tell it "remind me in 2 hours" and it creates the alarm. Muse parses intent and executes tasks locally — no third-party apps.

How it works

From download to first response in minutes.

No account. No email. No sign-up. Three steps and you're talking to your AI.

Install the app

Download Muse from the Play Store. No account, no email, no permissions asked upfront.

Pick a model

Choose from 1 GB (fast) to 8 GB (powerful) AI models. Downloaded once, stored locally, runs forever.

Start chatting

Type or speak. Muse responds in real-time using your phone's processor. Zero network calls.

Stay private

Conversations are encrypted locally. Export anytime. Delete and they're gone forever — we have no copy.

Performance

Real numbers. Real hardware.

These are measured values on typical mid-range Android devices. Smaller models are faster; larger models reason better. You pick the right balance for your device.

Inference Speed by Model

Tokens per second on a mid-range Android (Snapdragon 8 Gen 1, 8 GB RAM)

Gemma 3 1B reaches 22 t/s Phi-3 balances speed + quality

Speed vs. Context Length

How tokens/sec changes as conversation grows (Gemma 3 1B)

Speed stays above 12 t/s up to 1024 tokens

First-Token Latency: Muse vs Cloud AI

Time to first character of response (ms). Lower is better. Muse has zero network overhead.

Muse (on-device)

Cloud AI (fast connection)

Cloud AI (poor connection)

Muse: 150–400 ms (no network) Cloud: 500–5000 ms (depends on network) Airplane mode: Muse still works — cloud = 0

Privacy

Not private by policy. Private by physics.

Cloud AI companies promise not to train on your data — then update their terms. Muse can't make that trade-off because there's no server receiving your data in the first place.

of your conversations reach any server

Muse · All processing on-device · No telemetry on chat content

100%

of cloud AI conversations leave your device

ChatGPT, Gemini, Copilot · Every message sent over the internet

Data Sent to Cloud per Conversation

Percentage of conversation data that leaves your device (lower = more private)

Muse sends 0% of conversation data

Privacy Capability Comparison

Across five privacy dimensions — Muse vs cloud AI alternatives

Muse leads on every privacy dimension

TurboQuant · Memory

4× more context. Same RAM.

Muse uses TurboQuant KV cache compression (from ICLR 2026, Google Research) to dramatically reduce memory usage during inference — letting you have much longer conversations on the same device.

KV Cache Memory vs. Context Length

How much RAM the conversation memory uses as you chat longer (Gemma 3 1B). Lower means more room for actual model weights.

Off (FP16 — baseline)

Q8_0 — 2× compression

Q5_0 — 3.2× compression

Q4_0 — 4× compression

Q4 saves 108 MB at 2048-token context Q8 is near-lossless at 2× savings Freed memory = longer conversations

Max Context Length at 36 MB RAM Budget

How many tokens fit in the same memory footprint — more tokens means the model remembers more of your conversation

Q4 fits 4× more conversation history

Quality Preservation at Each Compression Level

Dot-product correlation (attention accuracy). Above 0.98 is effectively lossless in practice.

Q8 is 99.9% identical to uncompressed Even Q4 stays above 98.5% accuracy

Model Guide

Pick the right brain for your phone.

All models run locally via llama.cpp in GGUF format. Smaller = faster. Larger = smarter. Here's how they compare across the dimensions that matter on mobile.

Model Comparison — Speed, Quality, Memory, Reasoning

Normalised scores across four key dimensions. Hover to inspect each model's strengths.

Gemma 3 1B — fastest, great for daily chat Phi-3 Mini — best balance of speed + reasoning Llama 3 8B — highest quality on flagship devices

Download Size vs Inference Speed

Model size on disk vs tokens per second. Smaller models fit on more devices.

Minimum RAM Required

Models need free RAM to load. This shows the floor for each model to run smoothly.

3 GB RAM is enough to start

How it compares

No other mobile AI app does all of this.

There are excellent desktop LLM runners and cloud AI apps. Nothing else combines full offline operation, persistent memory, mesh inference, voice mode, and vision on Android.

Overall Capability Comparison

Muse vs leading alternatives across six key user-facing dimensions. Higher = better.

Muse is the only app that scores well in all six dimensions ChatGPT wins on model quality — but requires internet and collects data PocketPal is offline but lacks memory, mesh, actions, and voice