🧠
The engine
Welvet
A from-scratch AI engine in Go that trains and runs neural networks on ordinary hardware. No special chips. No cloud. The foundation everything else is built on.
Learn about Welvet →
OpenFluke
AI that runs on your device,
not someone else's cloud.
Tools for artificial intelligence that stay private, offline, and free on the computers and phones you already own. No accounts. No subscriptions. No data leaving your machine.
Who we are
An independent lab for private, on-device AI
Most AI lives in someone else's data center. You rent it, and everything you type gets sent away. OpenFluke builds the opposite: intelligence that stays on your own hardware, open source, and yours to inspect.
🔒
Works offline
Run AI with the internet switched off. No API keys, no sign-ups, and nothing you do gets uploaded.
💻
Runs anywhere
The same engine works on a laptop, a phone, or in a browser: Windows, macOS, Linux, iOS, and Android.
🔎
Open and honest
Free, open-source code with results you can verify. We publish proof that it works, not just claims.
What we build
Three parts, one idea
Welvet is the engine, Octo is the app you drive with, and w2a is the inspection that proves the engine is safe.
🐙
The app
Octo
Grab a popular model, get it ready in one step, then chat, generate images, or create speech, all running locally through Welvet.
See what Octo does →✅
The proof
w2a
The honesty check. Measures Welvet and publishes the results, so you never have to take our word for it.
How we prove it →
What it can do today
Chat, images, and speech on your machine
Download a model once with Octo and use it offline. Same tooling, several kinds of AI.
💬
Chat
Talk to a local language model for questions, drafting, and ideas. Completely private.
🎨
Images
Turn text into pictures with on-device models. Your prompts never leave the machine.
🔊
Speech
Turn written words into natural-sounding audio, generated on your own computer.
Proof it works
Numbers, not vibes
Every claim is backed by a public test harness. Here's the latest full pass of the engine.
137,039
test cells passed
326/326
suite cases green
34
numeric types
21
quant formats
3
hardware backends
64
book chapters
One end-to-end run: 0 gaps, 0 fails, timed across CPU-tiled, Plan 9 SIMD, and WebGPU. Full breakdown on w2a.
Our released app
SoulGlitch: local AI you can play with
A playful voxel world with an AI companion that lives entirely on your device. Chat, explore, experiment. Nothing phones home. Live now on the App Store, Google Play, and Linux.
Learn everything
The Welvet feature book
64 chapters covering every part of the engine, each with a small runnable example, plus the Octo benchmark and the full validation report.