01
Airplane & offline travel
Draft emails, summarize notes, and brainstorm when Wi‑Fi is a myth. Model stays on the handset.
Travelers · digital nomads
Available on Google Play · Android 7.0+ · arm64
AI that never leaves your phone.
Run compatible GGUF models directly on Android with llama.cpp. Chat offline without uploading your conversations or creating an account.
- On-device inference
- No account required
- Offline after model download
- Compatible GGUF models
Cloud AI asks you to trust a server.
LlamaBox lets you test the device.
Privacy is not a policy paragraph — it is the architecture. Inference has no LlamaBox server dependency, so you can switch the network off and check.
Public release
Already running on real Android phones.
LlamaBox is publicly available on Google Play. Install the app, download or import a compatible model, and start chatting offline. Device compatibility and performance testing continue.
Testing status
- Available on Google PlayInstall the public Android app. No tester invitation is required.
- Device compatibility testing continuesModel loading, offline generation and performance vary by phone and model.
- Public source release plannedThe app source is not yet public. An AGPL-3.0 release is planned; no date is announced.
Verified test data
Not yet published
Devices tested
Not yet published
Android versions
Not yet published
Models tested
Not yet published
Crash-free sessions
We publish these only once they are measured on real hardware. See tested devices for the data schema and methodology.
Use cases
Built for moments when the cloud is the risk.
Not another chatbot wrapper. A local reasoning engine for people who need control, continuity, and silence.
02
Journalism & sources
Work with sensitive quotes and drafts without routing text through a third-party model API.
Reporters · researchers
03
Student study buddy
Explain concepts, quiz yourself, rewrite notes — without selling your study history to an ad stack.
Students · lifelong learners
04
Vision on the go
Photograph a whiteboard, receipt, or diagram. Compatible multimodal models describe it locally on CPU.
Field work · makers
05
Privacy-first personal AI
Journaling, health questions, private plans — LlamaBox does not upload conversations for cloud-model training.
Anyone who values silence
06
Developer sandbox
Experiment with GGUF models, sampling knobs, and system prompts on real mobile hardware.
ML tinkerers · OSS
07
Low-connectivity regions
Useful AI where bandwidth is expensive or censored. Download once; think offline afterwards.
Emerging markets · field ops
08
Hands-free voice readback
Generate a reply, then listen with system TTS while you walk, cook, or drive (safely parked).
Multitaskers
Product
Everything for private, on-device AI.
A complete local-LLM toolkit — models, chat, vision, voice, and visibility into what the phone is doing.
Core
Offline inference after setup
Once a compatible model is stored on the device, generation does not require the network. Airplane mode is a feature, not a failure state.
Multimodal
On-device vision
Camera or gallery → local encode → answer. Requires a compatible multimodal model and a matching projector. Image encoder is forced to CPU so the UI stays alive.
Model management
Download curated GGUF builds or import your own. Q4_K_M is recommended for mobile speed/quality — not a universal compatibility promise.
Voice output
System TTS reads answers aloud for hands-free follow-through. Behaviour depends on the Android TTS engine you have selected.
System monitor
Live RAM / CPU insight while a model is loaded — know what your phone is carrying.
Engine
llama.cpp · React Native New Architecture
JSI/TurboModules for streaming tokens without legacy bridge tax. Zustand + SQLite + AsyncStorage — each store does one job well.
Scope
CPU by default; experimental acceleration may be available on supported devices
A compatibility trade-off, not a speed claim. Avoiding device-specific acceleration means fewer dependencies and broader Android support.
How it works
Three steps. Then silence.
From install to private chat without wiring your life to someone else’s compute cluster.
01
Get it on Google Play
Android 7.0+ arm64. Install from Google Play — package com.llamabox.
02
Load a GGUF model
In-app download or import. Start small (~0.5–1B Q4) on mid-range phones.
03
Chat offline
Optional airplane mode. Tokens generate on-device; history stays in local SQLite.
Verification
What “on-device” means in LlamaBox.
Specific, checkable statements instead of adjectives. Where a claim cannot yet be verified from public source, we say so.
The claims
- Generation
- Bundled local llama.cpp
- Model weights
- Stored on the device
- Chat history
- Local SQLite
- Account for local chat
- Not required
- Network for model download
- Required, user-initiated
- Airplane-mode generation
- Works once a model exists
- Usable model size
- Limited by device RAM
- Compute
- CPU by default
- Vision
- Needs matching projector
Don’t trust the claim. Verify the behavior.
- Download or import a compatible GGUF model.
- Start a conversation and confirm you get a response.
- Enable airplane mode.
- Continue generating — inference proceeds on-device.
- Inspect Android per-app network activity if you want independent confirmation.
Full detail: network behavior · architecture · privacy · llms.txt · tested devices
The visual guide library
A clearer start with local AI.
Read 10 short guides, preview the illustrated pages and keep the original PDFs. No signup needed.
Product UI
Private AI that lives on glass.
Explore models, device information and generation controls. Select a product preview to view it full size.
Performance · evidence
Measurements, not estimates.
Performance varies by phone, model, quantization, context and thermal state. Results appear only after a measured run is recorded.
Runtime today
- Prompt processing
- Not yet published
- Token generation
- Not yet published
- Model load
- Not yet published
- Peak memory
- Not yet published
- Context default
- 2048 (4096 vision)
- Compute
- CPU · 4 threads · NEON
CPU by default; experimental acceleration may be available on supported devices — a compatibility trade-off that avoids device-specific acceleration dependencies, not a claim of higher speed.
Stack
- App
- React Native 0.81
- Architecture
- Fabric + JSI
- Inference
- llama.rn · llama.cpp
- Models
- GGUF (Q4_K_M recommended)
- State
- Zustand + SQLite
- Package
- com.llamabox
- Target API
- 36 (min 24)
Deep dive: architecture docs · llms-full.txt for AI agents.
Compatibility evidence
This phone, this model, this result.
The most useful thing we can publish is a real measurement matrix. The component and schema exist now; rows appear as verified runs are recorded.
| Device | Android | RAM | Model | Quantization | Load time | Generation | Result |
|---|---|---|---|---|---|---|---|
|
No verified device measurements are published yet. Device benchmark runs are being collected on real hardware. See tested devices for the schema, methodology and how to contribute a run. |
|||||||
- Results vary by CPU, memory bandwidth, thermal conditions, context length and model architecture.
- A model fitting in storage does not guarantee it will fit in usable RAM.
- Q4_K_M is a recommendation, not universal compatibility.
- Vision requires a compatible multimodal model and a matching projector.
- LlamaBox supports compatible GGUF models handled by its bundled llama.cpp build, subject to model architecture, available memory and device limits.
Positioning
Why not just use ChatGPT?
Cloud assistants win on raw model size. LlamaBox wins when the data must not leave the room.
| Capability | LlamaBox | Typical cloud AI app |
|---|---|---|
| Inference location | On your phone | Vendor servers |
| Works offline | Yes, after a model is stored on the device | No |
| Account required | No | Usually yes |
| Conversation telemetry | No app analytics or conversation telemetry SDK | Policy-dependent |
| Independently auditable | Not yet — source release planned | Trust the vendor |
| Largest frontier models | Bounded by device RAM | Yes |
| Vision on-device | Supported with compatible model + mmproj | Commonly processed on provider servers |
Limitations
Local AI has real limits. We show them.
If a device cannot run a model well, we would rather you learn that here than after a 3 GB download.
Smaller models
Local models are far smaller than frontier cloud systems. Expect weaker reasoning, less world knowledge and more errors.
Device-specific speed
Token rate depends on your specific CPU and memory bandwidth. There is no single number we can promise.
RAM constraints
Available RAM caps usable model size. Android may reclaim memory and interrupt generation.
Storage requirements
GGUF files run from hundreds of megabytes to several gigabytes.
Battery consumption
Sustained CPU inference drains the battery noticeably faster than ordinary app use.
Thermal throttling
Long generations heat the device; the SoC will slow itself down to compensate.
Slow prompt processing
Older devices process long prompts slowly, which delays the first token.
Architecture differences
GGUF is a container. Support depends on whether the bundled llama.cpp build handles that specific architecture.
Vision projectors
Multimodal use needs a matching mmproj file alongside the base model. Mismatches fail to load.
CPU-default generation
The current build does not use GPU or NPU acceleration.
Model licenses
Models carry their own licenses and usage restrictions, separate from the LlamaBox app.
First-load time
Initial download plus first model load is the slowest part of the experience.
Context: performance data · architecture · tested devices
FAQ
Questions, answered cleanly.
Citable facts for humans — and for answer engines via schema + llms.txt.
What is LlamaBox?
LlamaBox is an Android app that runs compatible GGUF language models on your phone using llama.cpp. Generation happens on the device, so chat does not require a LlamaBox server and no account is needed for local chat. Multimodal vision is supported with compatible models and a matching projector. LlamaBox is publicly available on Google Play.
Does LlamaBox need internet?
Not for chat, once a compatible model is stored on the device. A connection is needed only when you explicitly download a model, or when you use this website or the support email.
Is my data private?
The Android app does not upload chats, prompts or generated responses, and contains no analytics or crash-reporting SDK. Prompts, conversations and images stay in local storage. Separately, support email does send what you type in it — see the privacy policy for the exact boundary between the app, the website and the form.
What models can I run?
Compatible GGUF models supported by the bundled llama.cpp build, subject to memory and architecture limits. Prefer Q4_K_M. Small models (e.g. Qwen2.5 0.5B, SmolLM2 360M) suit mid-range phones. Vision needs a base model plus matching mmproj (e.g. Gemma 3 4B Vision).
Can it analyze images?
Yes with a compatible vision model and matching projector. Photo or gallery → local encoder (CPU) → text answer. Keeps the UI responsive by avoiding graphics contention.
Does it use the GPU?
No. The current Android build performs CPU-default inference. This is a compatibility trade-off — it avoids device-specific acceleration dependencies so the app runs on a wider range of hardware — not a claim that it is faster than accelerated runtimes.
Is it free / open source?
The app is provided at no cost; models carry their own separate licenses. On source: the Android app source is not yet public. It is planned for public release under AGPL-3.0, with a separate commercial license available. “LlamaBox” is a reserved trademark. Commercial enquiries: work.aalhad@gmail.com.
What Android versions?
Android 7.0 (API 24)+, arm64-v8a. The app targets API 36. iOS is a future target.
How do I get access?
LlamaBox is publicly available on Google Play. Install LlamaBox, then download or import a compatible model. No tester invitation is required.
For teams and builders
Deploy private, on-device AI in products and offline workflows.
Commercial licensing, enterprise deployment and OEM integration for managed devices, regulated environments and low-connectivity operations.
Deployment & licensing
Partnerships
Corporate
Available on Google Play
Your next conversation doesn’t need a server.
Get it on Google Play. Bring a phone. Leave the cloud behind.