Nano
Released349M Dense
Trained on 100B tokens.
Ivonar Native ternary LLM family.
Built for extreme efficiency: 1.58-bit weights, Mamba-3 mixers and MLA attention.
48.4average of six benchmarks
8×smaller than FloatLM 390M
7.5×faster than its own FP16 path
Run Ivonar Nano on your own computer
A chat page, a terminal chat and an OpenAI-compatible API, on NVIDIA GPUs or the CPU.
curl -fsSL https://raw.githubusercontent.com/LuisOezdem/ivonar-inference/main/install.sh | sh
Four scales, one technical direction
Mini
In planning2.0B Dense
The first full production scale.
Medium
Coming laterA big step up in capability, at a size that stays practical.
High
Coming laterThe flagship: Ivonar at full scale, built for the most demanding work.
one square = 250M parameters
Ivonar Nano, and how it scores
Against five published models from 390M to 700M, each scored as its authors report it.
Weights and kernels under Apache 2.0.
Zero-shot benchmarks6 models · 6 tasks
| Model | Ivonar Nano | TriLM 390M | FloatLM 390M | QuantLM 4-bit | BitNet b1.58 | LLaMA 700M |
|---|---|---|---|---|---|---|
| Parameters | 349M | 390M | 390M | 390M | 700M | 700M |
| Training tokens | 100B | 300B | 300B | 300B | 100B | 100B |
| Weights | Ternary | Ternary | FP16 | 4-bit | Ternary | FP16 |
| ARC-Easy | 56.0 | 48.6 | 51.0 | 49.6 | 51.8 | 54.7 |
| ARC-Challenge | 24.8 | 21.2 | 21.3 | 21.3 | 21.4 | 23.0 |
| HellaSwag | 32.8 | 32.0 | 35.7 | 35.1 | 35.1 | 37.0 |
| PIQA | 64.1 | 65.0 | 68.4 | 68.3 | 68.1 | 68.9 |
| Winogrande | 51.9 | 52.2 | 51.8 | 53.7 | 55.2 | 54.8 |
| BoolQ | 60.8 | 55.1 | 54.7 | 50.8 | 58.2 | 60.0 |
| Average (6) | 48.4 | 45.7 | 47.2 | 46.5 | 48.3 | 49.7 |
Seven times faster, eight times smaller
About 1,430 tokens per second on an RTX 4060 Ti, from 95 MB of weights.
Inference speed, RTX 4060 Titokens / s
Our own measurement, one request at a time; BitNet weights from the 1bitLLM reproduction.
Nano decodes 950 tokens/s without CUDA graphs, and about 130 on a Ryzen 9 3900X CPU (24 without its kernels).
TriLM and BitNet run on TriRun, the fastest ternary GPU kernel we could run; QuantLM on Marlin; FloatLM in FP16.
Independent engineering, backed by Germany's AI ecosystem
Founded and led by Luis Oezdem, from architecture to training infrastructure and data pipelines.
Compute and ecosystem partners
hessian.AI
High-performance training compute
TU Darmstadt
Research and compute access
Berlin Startup Support
Early-stage growth programme
Real BitLinear training, hybrid sequence mixing
Weight footprintbits per parameter
FP1616.0
INT88.0
W1.581.58
W1.58A8 BitLinear Training
Ternary weights and 8-bit activations from the first training step, not quantised afterwards.
Mamba-3 and MLA Hybrid
Mamba-3 mixers for long context, MLA attention for precise recall: 18 and 6 blocks in Nano.
Packed 2-bit Inference
Weights packed at two bits, run by dedicated CUDA and CPU kernels.
A European ternary model, built to be used
Competitive at a tenth of the budget and a seventh of the size
Language and code capability without scaling up dense parameters.
Minimal memory footprint
Ternary weights and Mamba-3 mixers keep memory and compute low.
Built for everyone
API access, subscriptions, and enterprise licences for teams that keep their data in house.
Talk to the team building it
Questions about the architecture, compute partnerships, or running Ivonar Nano can be sent directly.