Intel / article

Why Nextcloud bundled Olmo 3 for local, auditable AI

Source: allenai.org
Author: Ai2
Date: Published Nov 2, 2026

Context

Nextcloud, an open-source collaboration platform whose users include Germany's federal government and CSIC, Spain's largest public research organization, bundles Ai2's Olmo 3 with llm2, the local backend that runs AI features in Nextcloud Assistant, Mail and other apps, according to an Ai2 blog post. Olmo 3 holds green, the top grade in Nextcloud's Ethical AI rating. That grade requires three things: open-source software for both training and running the model, a trained model users can host themselves, and publicly available training data.

The training-data condition is what goes beyond running AI on your own servers. Many Nextcloud customers choose it to reduce their dependence on big tech, and Nextcloud software engineer Marcel Klehr says being able to inspect and reproduce how a model was built is vital to that goal. A model with downloadable weights but private training data can still run locally, but it would not meet that condition, because outsiders cannot see what it was built from. “As long as parts are proprietary, there is no full control over your data and AI results,” Klehr says.

Size was the other requirement. Nextcloud did not want to assume its users have 16 or 32 GB of graphics memory available, and Klehr says the 7-billion-parameter version gives a large audience a good experience out of the box. Administrators configure Olmo once through llm2, and the same self-hosted model then serves staff inside the Nextcloud tools they already use. Uses include asking about facts buried in documents, summarizing email threads, summarizing video calls after OpenAI's Whisper transcribes them, and reminders to follow up on unanswered emails.

The post is Ai2's own account, built on Klehr's quotes, and it gives no figures for how many people use Olmo through Nextcloud. What it does establish is a specific reason organizations pick a fully open model. Public institutions that must answer for their AI need to check how a model was made as well as where it runs. Klehr names stronger and support as the capabilities Nextcloud wants next.

Key takeaways · AI-distilled

Nextcloud's green Ethical AI rating requires open training and code, self-hostable weights, and public training data; Olmo 3 earned it, and Nextcloud bundles the model with llm2, its local backend.
Nextcloud picked the 7B Olmo 3 partly for size: engineer Marcel Klehr says the company does not want to assume users have 16 or 32 GB of VRAM, and the 7B model gives a broad audience a working local setup.
Because Nextcloud exposes models through a shared text-processing layer, administrators deploy Olmo once and staff use it inside Assistant, Mail, and other apps, for example to summarize email threads and transcribed calls.

Terms in this piece · Glossary

tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source allenai.org