Noeum.ai Releases Noeum-1-Nano, a From-Scratch Mixture-of-Experts Model Trained on 18B Tokens

Independent Austrian AI lab demonstrates data-efficient reasoning performance in nano-class category and outlines roadmap toward multimodal, multilingual scaling

Vienna, Austria - [Month Day, 2026] - Noeum.ai, an independent AI research and engineering lab based in Austria, today announced the release of Noeum-1-Nano, a nano-scale Mixture-of-Experts (MoE) language model featuring 0.6B total parameters (~0.2B active). The model was trained entirely from scratch using 18 billion tokens-approximately 20,667× less training data than many standard models in its class-demonstrating a data-efficiency thesis focused on extracting stronger reasoning per unit of compute.

Unlike many small-model releases that rely on inherited pretrained weights, Noeum.ai states that Noeum-1-Nano was built without pretrained checkpoints, aiming to validate an efficiency-first training and architecture stack under constrained resources.

In reported evaluations, Noeum.ai notes that benchmark comparisons were conducted with the model's optional "thinking mode" disabled, with results showing above-average nano-class performance, including a #1 ranking on MRPC (semantic equivalence) among comparable models.

"Noeum-1-Nano is a proof-point for an efficiency-first approach," said Bledar Ramo, Founder of Noeum.ai (http://Noeum.ai). "By iterating quickly at the nano scale, we validate what truly improves reasoning and reliability before committing resources to larger-scale training. The goal is to compete through architectural and data efficiency-not brute-force scale."

● System-2 'Think Mode' for Verification and Self-Correction

Noeum-1-Nano includes an optional System-2 style "think mode", designed to support multi-step verification and self-correction. In demonstrations provided by Noeum.ai, enabling this mode improves performance on tasks requiring intermediate reasoning steps (e.g., applying formulas such as distance = speed × time) and reduces failures seen in standard generation.

● Roadmap: Scale Only What's Proven

Noeum.ai's roadmap focuses on scaling validated techniques into a realistic-sized model with multimodality and multilingual support, trained on 1-3 trillion tokens. The lab cites ongoing work on long-context efficiency and self-correcting reasoning pipelines, with the broader objective of advancing controllable reasoning capabilities while maintaining computational and data efficiency.

● Next Milestones

Noeum.ai highlighted several near-term milestones of interest to researchers and industry partners:

Publication of reproducibility materials (evaluation configurations, scripts, reruns)
Intermediate-scale checkpoints to validate that nano-scale gains persist at a larger scale
Clear product pathways in efficiency-sensitive environments (e.g., on-prem, edge, and industrial deployments)

● Availability

Model details, benchmark tables, and technical documentation are available via the public model card and Noeum.ai's official website (http://Noeum.ai).

● About Noeum.ai

Noeum.ai is an independent AI research and engineering lab based in Austria, building next-generation intelligent systems with an emphasis on data efficiency, architectural innovation, and reproducible evaluation. Noeum.ai develops models end-to-end-from pre-training through alignment and benchmarking-entirely in-house.

The lab's efficiency-first approach aligns with emerging European initiatives to expand AI compute access while maintaining competitive development costs.

Media Contact
Noeum.ai
Email: contact@noeum.ai
Website: noeum.ai

This release was published on openPR.