AI Tech v0.9.2
A conversational AI assistant 100 % on your phone. No Internet connection, no cloud, no subscription.
Security audit — May 2026Why AI Tech
An AI assistant that shares nothing with anyone. Really.
Zero Internet
No INTERNET permission in the manifest. The app is technically incapable of sending any data.
Encrypted chats
AES-256-GCM with a unique key in the Android Keystore. Session-bound AAD: a moved file is unreadable.
Panic mode
Wipes key + history + models + settings in one tap. Final, unrecoverable.
Bilingual FR / EN
Full interface in French and English (since v0.6.0). Language selector in Settings, follows the system locale by default.
Open models
Gemma 4 E2B / E4B (new in v0.8), Gemma 3, Qwen 2.5, Phi-4, Llama 3, DeepSeek-R1. Automatic prompt-format detection.
Speculative decoding (MTP)
Automatically enabled on Gemma 4 when the `.litertlm` contains a draft model. ~30 % boost on tokens/s throughput with no quality loss.
Readable formulas
Automatic LaTeX → Unicode conversion at render time: H₂O, E = mc², √x, α, subscripts and superscripts. No raw `$\text{...}$` on screen.
Token-by-token streaming
Response shown as it is generated, like a real chat. Instant cancellation.
Full settings
Temperature, top-K, context length, multi-model management. You stay in control.
Screenshots blocked
FLAG_SECURE active on the chat screen. No screenshot, no preview in Recents.
MediaPipe format
Compatible with .task and .litertlm. Magic-bytes check on load.
Quick prompts
Improve a text, translate, summarize, explain, rephrase, brainstorm — one tap.
Export
Share a conversation as Markdown, with an explicit warning before sending.
Free forever
No subscription, no quota, no freemium. Apache 2.0, truly free.
Screenshots
How it works
AI Tech runs an open-source large language model directly on your phone, via Google's MediaPipe LLM Inference library.
1. You download a model
From Kaggle or HuggingFace, in .task or .litertlm format. The recommended model to get started is Gemma 3 1B int4 (554 MB) — excellent in French, very fast.
2. You import it into the app
The onboarding screen guides you to the Files Tech custom picker which validates the format before import (extension, minimum size 50 MB, magic-bytes). The model stays on your storage : the app does not copy it.
3. You chat
Multi-turn conversation, the model remembers context. History is encrypted on disk after each completed message : if the OS kills the app, nothing is lost. Panic mode wipes everything in one second.
Measured performance
On Gemma 3 1B int4, the recommended model.
Galaxy S24 FE (2024)
43 tok/s · first-token 253 ms
Above cloud ChatGPT, offline, on 554 MB.
Galaxy S9 (2018)
9.8 tok/s · first-token 3.3 s
Comfortable human reading speed. Any Android 2018+ with ≥ 4 GB RAM works.
Xiaomi POCO C75 (2024)
7.7 tok/s · first-token 2.0 s
Modern entry-level (Helio G81). Usable, pleasant streaming reading.
AI Tech vs ChatGPT, Claude, Gemini
Not smarter. More respectful.
| Criterion | AI Tech | ChatGPT / Claude / Gemini |
|---|---|---|
| Internet required | No, ever | Always |
| Data sent | None | All of it |
| Long-term cost | Free forever | Recurring monthly subscription |
| Quotas | None | Limited |
| Server outages | No effect | App unusable |
| Airplane mode / abroad without data | Works | Unusable |
| Behavioral tracking | None | User profile |
| Verifiable source code | Apache 2.0 | Black box |
| Real panic mode | 1-sec wipe | Nonexistent |
Who it is for
- People who care about their privacy (post-Snowden, post-GDPR)
- Professionals handling sensitive data (lawyers, doctors, therapists, journalists)
- Travelers and nomads without reliable data connection
- Those who refuse yet another €20 /month subscription
- Anyone who wants to understand what their tool does — and verify it
Honestly, the limits
- A 1 to 2 billion parameter model remains weaker than GPT-4 or Claude Sonnet (10,000× bigger) on complex questions.
- No web search, no image generation, no vision (at this stage).
- Inference uses the CPU — extended use : keep a charger nearby.
Open source code
For a local AI assistant, independent code audit is essential — it is the only proof that privacy promises hold. AI Tech is published under the Apache 2.0 license, you can read, verify and build it yourself.