mnemo
It knows nothing.
Random weights. No vocabulary. Mnemo builds its own tokenizer from the first text it reads, then starts guessing the next word, badly.
parameters 12,041,216
Raise an AI from nothing.
Mnemo grows a brand-new language model on your own computer. No borrowed brains, no cloud. You choose how big it starts and how much it can remember. It reads, it grows, it learns to talk.
$
Paste into Terminal. Takes a few minutes. Then type mnemo.
scroll to watch it grow ↓
mnemo · day 0
params12M
loss10.37
read0 tokens
Shapes of language.
After a few minutes of reading, the gibberish gets punctuation, real words, the rhythm of sentences.
It learns to talk.
One click runs thousands of everyday conversations. Eight minutes later it says hi back, tells jokes, and knows its own name.
Then it gets bigger.
When it has learned enough for its size, Mnemo grows wider and deeper without forgetting anything: 12M → 99M → 545M → 1B+, as far as your RAM allows.
🌐 General💻 Code📚 Literature🔬 Science➗ Math
📰 News💬 Conversation✨ Your own mode
Everything a growing
mind needs.
design
You pick the brain.
Starting size, growth ceiling, context up to 128K tokens, training window, focus. A setup wizard shows memory use and learning time before it's born.
web
Reads the internet
/learn coding finds the best pages and studies them. It respects robots.txt and stays polite.
browser
Drives its own browser
Click, type, scroll. Every action becomes training data, so it slowly learns to browse.
memory
Never forgets
Everything it reads is searchable and pulled into context when you chat. Remove anything anytime.
files
Learns your stuff
Drop in PDFs, Word docs, ebooks, whole code folders. It remembers and trains on them.
private
Lives on your machine.
The model, memory, chats and files stay in ~/.mnemo. Runs on Apple Silicon, NVIDIA GPUs, or plain CPU.
macOSWindowsLinux
studio
Images, video and 3D.
Images, video and 3D.
On your hardware.
Download open generative models from a curated catalog and run them locally. Every model shows its license up front, so your team knows what's cleared for commercial work.
image
Segmind SSD-1B1024px · Apache-2.0 · commercial
DreamShaper 8OpenRAIL-M · commercial
SDXL Turbo1–4 steps · research only
video
Wan 2.1 1.3B480p · Apache-2.0 · commercial
CogVideoX 2B720×480 · Apache-2.0
ModelScope T2V8 GB GPUs · research only
3d
Shap-Etext → .glb / .obj / .ply · MIT
More soonimage → 3D, textures
Tuned for your machine.
NVIDIAAny CUDA GPU. Pick the card, budget against its VRAM, multi-GPU boxes supported.
Apple SiliconMetal acceleration on M1 through M5, unified memory used efficiently.
CPUNo GPU? Choose thread count and a small model; it still learns.
Every knobWidth, layers, context to 128K, precision, optimizer, batch, learning rate, memory cap.
plugins
Teach it new
Teach it new
tricks.
A few lines of Python in ~/.mnemo/plugins/ become a slash command. Ships with
/wiki, /define, /calc, /time and /remember.
Honest answers.
How smart will it get?
It starts out writing gibberish. After the conversation course it handles small talk. With days of reading it writes real sentences about what it studied. It won't match ChatGPT or Claude, which train on thousands of GPUs. Mnemo is about owning and watching the whole thing.
Why a command instead of a download?
Downloaded apps need a paid Apple certificate, or macOS calls them "damaged". The installer builds Mnemo on your own computer, so it just works.
What do I need?
Mac with Apple chip (8 GB works, 16 GB+ is better), Windows 10/11, or Linux. An NVIDIA GPU makes it much faster. Keep 5–20 GB free for its reading library.
Why do I need an account?
It keeps Mnemo yours across updates and is how future features (sharing models, cloud training) will work. Your model and data still never leave your computer.
Is it free?
Yes, free to use. The source is public so you can see exactly what runs on your machine, but it may not be redistributed or sold. Pro extras may come later. Sign in and join the waitlist.