How Olmo 3 earned the highest ethical rating from Nextcloud

For Nextcloud, an open-source collaboration platform used by organizations including Germany’s federal government and CSIC, Spain’s largest public research organization, running AI on infrastructure an organization manages rather than through a third-party cloud is only one part of digital sovereignty.

Nextcloud also needed a model that met public organizations’ requirements for transparency and accountability. Olmo 3 earned the company’s highest Ethical AI rating: green. That rating requires open-source training and inference software, a trained model users can self-host, and publicly available training data.

Many customers choose Nextcloud specifically to reduce their dependence on big tech, so the company wants the AI models it supports to offer the same kind of independence and transparency: users should be able to know what a model is built from, reproduce it, and understand how it was made.

“We deeply believe in the four freedoms of free software and are convinced that being able to inspect and reproduce how the model was built is vital,” says Marcel Klehr, a software engineer at Nextcloud.

That philosophy shaped the models Nextcloud supports through llm2, the local LLM backend that powers AI features in Nextcloud Assistant, Mail, and other apps. Depending on the model, users can generate and improve text, create headlines, and extract topics—all while running the model on hardware their organization controls.

Olmo 3 is one of the models Nextcloud bundles with llm2. “We do want to offer models to our users that fulfill this goal,” Klehr says. “Olmo 3 was a natural choice for this reason.”

A fully open model that fits local hardware

Olmo’s openness was only one factor in Nextcloud’s decision to adopt the model for its customers. The company also needed a model that could run well on consumer hardware — one efficient enough to deliver a good local experience without requiring high-end GPUs.

“We don’t want to assume that our users have 16 or even 32 GB of VRAM readily available,” Klehr says. “The 7B model makes sure that a large audience gets a great out-of-the-box experience.”

For an organization, Olmo can be configured centrally through llm2 and then made available to employees through the Nextcloud tools they already use. Staff don’t need to connect to Olmo separately or move their work into a standalone chatbot; they can invoke the model from Nextcloud Assistant or directly within supported workflows.

Klehr says LLMs like Olmo let users delegate repetitive or tedious tasks to the Assistant’s agent and ask questions about facts buried in their documents. Other uses include summarizing incoming email threads and long video calls after Whisper, OpenAI’s speech-to-text model, has transcribed them, and reminding users to follow up on unanswered outgoing emails.

Because Nextcloud exposes language models through a common text-processing layer, administrators can deploy Olmo once and extend the same fully open, self-hosted model across multiple applications and users.

“Fully open models like Olmo create the possibility for an AI that can be fully transparent, explainable, auditable, and controllable,” Klehr says. “As long as parts are proprietary, there is no full control over your data and AI results.”

For customers balancing privacy and local compute

Those benefits map closely to the needs of many organizations Nextcloud already serves. Its clients include public-sector agencies, research institutions, and others that work with data they may not want to send to an outside AI provider. CSIC, for example, uses Nextcloud to provide roughly 17,000 researchers with a self-hosted environment for sharing and working with research data.

For organizations that also need to keep the infrastructure demands of local AI manageable, Olmo 3 7B has another practical advantage—its relatively small size. Smaller models generally require less compute than much larger ones, making it easier to keep AI workloads on infrastructure an organization already operates without taking on the cost and hardware requirements of a substantially larger model.

“As our customers chose Nextcloud to become independent from big tech, we believe that a larger percentage of our customers would use AI if it were to be a model that is fully open, like Olmo, and equally performant,” Klehr says.

“And this is only the start. We’re working on extending our AI features dramatically,” Klehr says.

The capabilities Klehr would like to see next – including stronger tool use and multimodal support – also align with areas Ai2 is investing in. We’re developing models that can work across more types of data and act more autonomously, including by planning and using tools.

As that work carries forward into future fully open models, it could expand the range of Nextcloud workflows where organizations can get those capabilities without giving up the transparency that drew Nextcloud to Olmo in the first place.