Mycel builds AI for science, research and innovation. We develop specialised models that reason over evidence, coordinate tools and support the work behind discovery, with people in control.

Like mycelia networks that connect a forest underground, we build systems that link what matters. Mycel is rooted in Germany, and its work spans research institutions, enterprises and cities.

How we release models

• Open weights. They run on hardware you own.
• Full evaluations. Each release documents how it was tested.
• Plain limitations. We say what a model cannot do.
• Citable. Every release gets a DOI.

Models

Numo-1-E2B (released)

Numo-1-E2B is a fast decision model for scientific literature. From a paper's title and abstract, it decides what kind of study the paper is, which field it belongs to, or whether it is relevant to another paper, with a calibrated probability for every answer.

• Accuracy: 94.3% on 600 PubMed papers labelled by NLM indexers.
• Calibration: the lowest calibration error of every model we tested.
• Speed: about 124 ms per decision on one RTX 3060, fully offline.
• Licence: Apache-2.0, built on Google's Gemma 4 E2B.
• DOI: 10.57967/hf/11188.

Ark-1 (research in progress)

Ark-1 is the first model in the Ark series. It is a research project developing specialised AI that reasons over evidence, coordinates tools and supports research and innovation workflows, starting with research funding.

• Now: post-training on MareNostrum 5 at the Barcelona Supercomputing Center.
• Next: Ark-1.5 is planned with the HammerHAI AI Factory.
• Release: public releases will follow evaluation and review.

This is ongoing research. It makes no claim of a released model or of validated performance. The infrastructure acknowledgements describe computing access only and do not imply endorsement.

Products and initiatives

• ResearchArk is our agentic operating system for research, innovation and funding. It helps you discover funding, develop proposals, build consortia and explore the research graph. It is live and free to start. Its code is on GitHub.
• Agentpower is an open, public-domain framework. It measures the cognitive work output of human–AI systems.
• More initiatives: open frameworks, free and open-source tools, and open-access resources.

Collaborate

We welcome conversations with researchers, institutions and partners working on reliable AI for research and innovation.

mycel-ai.de · dev@mycel-ai.de · GitHub · LinkedIn · Mycel UG (haftungsbeschränkt), Berlin, Germany