Noeum

Genuine Reasoning. Architectural Intelligence.

About Noeum

Noeum is an independent AI research & engineering lab based in Austria 🇦🇹. We are building the next generation of intelligent systems by validating high-efficiency scaling hypotheses.

The name Noeum derives from "mind," "meaning," and *"thought"*—reflecting our mission to construct genuine reasoning systems rather than models that rely solely on statistical mimicry. We are one of the few labs in Europe executing the full AI pipeline—from architecture design and pre-training to alignment—entirely in-house.

Our Methodology

"Iterate fast at nano-scale; scale only what works."

We challenge the industry standard of "brute-force" scaling. Instead, we focus on architectural intelligence and data density.

Data Efficiency: Proving that high-signal data can rival trillion-token datasets.
From Scratch: No pretrained weights, no shortcuts. Custom tokenizer, pre-training, SFT, GRPO, and DPO.
Innovation: Integrating Recursive Reasoning, Agentic Data Synthesis, and Hyper-Efficient Architectures.

Featured Model: Noeum-1-Nano

Our proof-of-concept model demonstrates that innovative techniques can rival brute-force scale.

Noeum-1-Nano-Base
- Specs: 0.6B MoE (0.2B Active)
- Efficiency: Trained on only 18 Billion tokens (a 20x to 667x data reduction compared to standard models like Qwen2 or TinyLlama).
- Performance: Matches or exceeds widely-used 1B models, securing top ranks on logic-heavy benchmarks like BoolQ and MRPC.

The Vision

With our nano-scale hypotheses validated, we are preparing to apply these techniques—including multimodality and recursive reasoning architectures—to realistic-sized models trained on 1 Trillion+ tokens. We believe this architectural leap allows independent research to compete directly with the largest AI labs.