Nano
Dense350M
In trainingA 350M dense tier for fast iteration and end-to-end validation. It uses 18 Mamba-3 blocks plus 6 MLA blocks, all parameters active, and no MoE routing.
Active350M
Base ctx4k
Long ctx8k
Native ternary LLM family · Germany
IvonarNative ternaryLLM family.
A native 1.58-bit LLM family built for extreme efficiency. Powered by W1.58A8 BitLinear architecture, Mamba-3 state-space mixers, and MLA attention, scaling seamlessly from dense tiers to sparse MoE across four model scales.
01 · About
Four scales, one technical direction
Ivonar is a native ternary LLM family developed independently in Germany. It spans four training scales: Nano, Mini, Medium, and High. Each tier shares the same technical direction while scaling model capacity, active parameters, context length, and dense-to-sparse architecture choices.
Mini
Dense1.5B
A compact 1.5B dense model for the first full production scale. It keeps all parameters active, combines Mamba-3 with MLA, and avoids expert routing.
Active1.5B
Base ctx8k
Long ctx64k
Medium
Sparse8.0B
An 8.0B total parameter tier with 3.0B active parameters. Sparse capacity is limited to feedforward blocks with Top-2 routing; attention and sequence mixers stay shared.
Active3.0B
Base ctx16k
Long ctx128k
High
Sparse32.1B
A 32.1B total parameter flagship tier with 6.3B active parameters. It extends the FFN-only sparse MoE pattern with Top-2 routing and the longest context target.
Active6.3B
Base ctx32k
Long ctx256k
02 · Scaling sheet
The family, tabulated
The same architecture read as a specification: total and active parameters, base and long context targets, and where sparse expert routing enters the stack.
Ivonar model family
W1.58A84 tiers2026
| No. | Tier | Total | Active | Base ctx | Long ctx | Routing |
|---|---|---|---|---|---|---|
| 01 | Nano | 350M | 350M | 4k | 8k | Dense · no MoE |
| 02 | Mini | 1.5B | 1.5B | 8k | 64k | Dense · no MoE |
| 03 | Medium | 8.0B | 3.0B | 16k | 128k | FFN-only MoE, Top-2 |
| 04 | High | 32.1B | 6.3B | 32k | 256k | FFN-only MoE, Top-2 |
Context values are per-tier training targets; long context is reached in a separate continuation phase with LongRoPE2.
03 · Team
Independent engineering, strengthened by Germany's AI ecosystem
Ivonar was founded and is led by Luis Oezdem as an independent research and engineering initiative focused on efficient ternary LLM development, combining architecture, training infrastructure, and data pipelines under a unified vision.
Our execution model is simple: validate performance on Ivonar Nano, then scale carefully into Mini, Medium, and High with clear quality and efficiency targets.
Compute & ecosystem partners
hessian.AI
High-performance training compute
TU Darmstadt
Research and compute access
Berlin Startup Support
Early-stage growth programme
04 · Architecture
Real BitLinear training, hybrid sequence mixing
The ternary stack is built around real W1.58A8 BitLinear training in PyTorch: ternary-weight STE, INT8-activation STE, Mamba-3 state-space mixers, MLA latent attention, and FFN-only sparse routing where the larger tiers need it. Packed or custom kernels are future inference optimizations, not the current training GEMM.
18 / 6 Mamba-3 · MLA, NanoTop-2 FFN routing, Medium & High
W1.58A8 BitLinear Training
Ivonar trains real BitLinear models with ternary-weight STE and INT8-activation STE in PyTorch. It is not yet a custom bit-packed CUDA training GEMM; packed/custom kernels are future inference work.
Mamba-3 and MLA Hybrid
Mamba-3 state-space mixers handle efficient long-range sequence processing. MLA latent attention provides retrieval-focused attention, with Nano explicitly using 18 Mamba-3 blocks and 6 MLA blocks.
Dense and Sparse Tiers
Nano and Mini are dense models with no MoE. Medium and High use FFN-only sparse MoE with Top-2 routing, keeping sparse capacity out of attention and sequence-mixing blocks.
Long Context with LongRoPE2
Context targets scale from 4k base and 8k long in Nano up to 32k base and 256k long in High, using separate base and long-context continuation phases.
05 · Efficiency
A powerful European ternary model built for efficient use
Ivonar aims to become the leading European ternary LLM: real W1.58A8 BitLinear training in PyTorch today, honest dense and sparse tiering, and practical deployment work that can later benefit from packed or custom inference kernels.
Strong results with compact models
Ternary BitNet layers achieve competitive language, code, and math capability without dense-parameter scaling. The goal is practical quality across daily work, prototypes, and research tasks.
Individuals, builders, research workflows
Minimal memory footprint
Ternary weights, Mamba-3 mixers, and dense-to-sparse scaling keep active compute controlled. Nano and Mini are dense; Medium and High use FFN-only sparse MoE with Top-2 routing.
Practical inference and controlled scaling
Built for everyone.
Ivonar serves both B2C and B2B markets. Our monetization strategy relies on commercial API access for developers, premium subscriptions for end-users, and flexible enterprise licensing for businesses that require strict data privacy and local deployment.
Flexible access for users, builders, and teams
06 · Roadmap
From Nano validation to sparse scale-up
Architecture defined. Nano training is active. Mini, Medium, and High scale from there.
Architecture
Completed
Hybrid ternary architecture with PyTorch W1.58A8 BitLinear training, Mamba-3 mixers, MLA attention, dense Nano/Mini tiers, and FFN-only sparse MoE for Medium/High is defined.
Nano Training
In Training
Training the 350M Nano tier with 4k base and 8k long context for fast iteration, pipeline validation, and end-to-end BitLinear quality checks.
Mini Training
Upcoming
Next dense tier: 1.5B total and active parameters, 8k base context, 64k long context, Mamba-3 + MLA, and no MoE routing.
Sparse Scale-Up
Upcoming
Medium and High add FFN-only sparse MoE with Top-2 routing: 8.0B total / 3.0B active for Medium, then 32.1B total / 6.3B active for High.
07 · Contact
Talk to the team building it
Questions about the architecture, compute partnerships, or early access to the Nano tier can be sent directly.