Vesper
A family of AI models that Eyad designed and trained from zero, on one home PC.
One graphics card. Billions of words.
No borrowed brain. Every weight learned here.
Reading
First it reads billions of words from the web, Wikipedia, textbooks and Python code, learning one thing: guess the next word. Do that billions of times and grammar, facts and patterns fall out.
15B tokens for Motif 2 · 12B for Sonata 1Practice
Then it practises hundreds of thousands of example conversations, worked math problems and reading tests, so it answers you instead of just continuing your text.
conversations · worked problems · reading testsTools
It learned to use a calculator, look things up on Wikipedia and show its sources, check the weather, and make and edit files. Python it writes is run in a sandbox before you see it.
calculator · Wikipedia · files · Python sandboxTesting
A new model only ships if it beats the old one: a 106-question release test, five standard benchmarks, and a check that test questions didn't leak into its training data.
Sonata 1: 87 of 106 · beat Motif 1.1 on 5 of 5 benchmarksNamed like music. The bigger the piece, the bigger the model.
Before these came the Storytellers (45–98M), which could only write simple children's stories, then Motif 1, 1.1 and 1.2.
Smaller models, each built for one job.
~110M · COMINGTwo voices in conversation. Built only for role-play: game characters that stay in character, cheap enough for small indie games.
50M · COMINGThe quick one. Short lines and reactions for games, fast enough to feel instant. It launches together with Duet.
PLANNEDVoices for characters, so a game gets the line and hears it spoken. The game pack launches with an open voice model; Timbre is our own, coming later.
PREVIEWA coding lab that writes, runs and fixes Python in your browser. The preview runs on Sonata 1; its own code-trained model is next in line.
One piece at a time.
- Sonata 1.5TRAININGA 16,000-token memory, and better at following the instructions you give it.
- The game packDuet 1 and Collab 1, released together, with character voices.
- Motif 2.1Longer memory and better conversation practice for the fast model.
- Vesper Code 1A model of its own for the coding lab, trained on billions of tokens of code.
- Timbre 1Our own voice model.
- Symphony 1The first 1.2-billion-parameter model, on rented GPUs.
After that: Sonata 2, Motif 3, Duet 2 and bigger Symphonies. Plans can change; the release test decides.
Build with Vesper
Put Vesper in your own apps, bots and games. The API works like OpenAI's, so most existing code needs two changes. Built for games and hobby projects: from $0.05 per million tokens, free credit to start, and ready-made examples for Godot and Unity.
Get an API key →model: sonata-1
"Hi! I'm Vesper. What are we building?"
■ 200 OK · 14 tokens
Small, and it says so.
A model this size often gets facts wrong, sometimes confidently. Check anything important.
A confidence label under each reply, the Wikipedia articles it used, and whether its code actually ran.
The same computer that trains the next model also answers you, so when it's busy or off, Vesper is slower or offline.
They're saved in your browser, not on the server. See the privacy page.