Specialists,
trained in threes.
The Grimm Series is our model family, and each name in it is a class, not a single model. A class is a size and a home: how big the model is, and where it runs. Inside each class live specialists, models trained for one field each, so a small model can be genuinely good at its job instead of mediocre at everything.
Specialists are fields, not features: the core subjects that push a field forward, math, physics, history, religion, plus the computing disciplines. Every specialist is trained at every class scale: the same mind, from your phone to your workstation. All of them are open mixture-of-experts (MoE) models with three-value weights {−1, 0, +1}, the specialists are the experts, and the model routes each job to the mind trained for it.
⬢ STATUS // In development. No Grimm model is released or downloadable yet, the first proof of concept is being trained now (below).
Thimble
The pocket class, and the only one that lives on your phone. One to four billion parameters: small enough to ride in a pocket, large enough to be a real specialist. The always-with-you members of the family.
Cinder
The heavyweight class. Twenty-seven billion parameters, too heavy for a phone, at home on a desktop or workstation you own. When the pocket model shouldn't decide alone, the job climbs here.
Raven
The heaviest class. The top of the escalation ladder, our largest planned family of specialists. It stays unlisted on purpose: no specs, no numbers, until there is something real to measure.
Each class fills with specialists, one model per field, trained deep instead of broad. The program starts where the lab's research is deepest, then works outward through the core subjects:
⬢ EVERY CLASS // Each specialist ships at every class scale, the same field, from a Thimble on your phone to a Cinder on your workstation and beyond, each with a measured scorecard, or not at all.
⬢ ROADMAP // Proof of concept first, at the smallest scale that proves the method. Then the specialists fill out across the classes. Bigger means smarter and slower, the app will help you pick.
EFFICIENT BY DESIGN効率設計
Every weight, the numbers a model is made of, is stored as −1, 0 or +1 instead of a long decimal. About 1.6 bits per weight instead of 16: roughly a tenth of the memory, with multiplication swapped for simple addition. MoE routing means each question wakes only the experts it needs, not the whole library. Our edge is the tuning data and the judging harness, not the weights. Those ship open.
WHY TERNARYなぜ三進法か
⬢ STRAIGHT TALK // This is ternary weight compression running on standard hardware, less memory traffic, no multiplication, not ternary circuits. We say so because we'd want to be told.
⬢ FURTHER READING // The model program's documents, the method, the class dossiers, the theory, are in review while the theory is being tested, and publish to docs once confirmed. Until then, this page is the program's current, honest status.