Ivonar

A native ternary language-model family, developed independently in Germany. Every large weight matrix holds only −1, 0 and +1, trained in that form from random initialization rather than quantised afterwards, and shipped packed at about two bits per parameter.

Ivonar Nano is the first finished model: 349M parameters, 100B training tokens, 94.6 MB on disk, 1,078 tokens per second on an RTX 4060 Ti. Apache-2.0.

Mini (1.5B), Medium (8.0B) and High (32.1B) are planned; the architecture and the training pipeline carry across all four.

• Model: Ivonar Nano
• Runtime: github.com/LuisOezdem/ivonar-inference
• Project: ivonar.com