Olmo 3 won Nextcloud's top green ethics rating and now runs on its customers' own machines

Nextcloud, an open-source collaboration platform, gave Ai2's Olmo 3 its highest Ethical AI rating: green. Ai2 reports this in a post on its own blog, with comments from a Nextcloud engineer.

What green requires

The rating asks for three things. The training and inference software must be open source. Users must be able to self-host the trained model. And the training data must be public.

Among Nextcloud's customers are the German federal government and CSIC, the Spanish national research council, which is the biggest public research body in Spain. Many of them chose Nextcloud to depend less on big tech, so the company wants the same independence from its AI models.

Small enough for ordinary hardware

Openness was only one reason for the choice. Nextcloud also wanted a model that runs well on consumer hardware.

Marcel Klehr, a Nextcloud software engineer, put it this way: "We don't want to assume that our users have 16 or even 32 GB of VRAM readily available." VRAM is the memory on a graphics card. The 7B model, he says, gives a large audience a good out-of-the-box experience.

Olmo 3 is bundled with llm2, the local backend behind AI features in Nextcloud Assistant, Mail and other apps. An administrator sets it up once. Staff then use it inside the tools they already have, with no separate chatbot.

What people use it for

Depending on the model, users can generate and improve text, write headlines and extract topics. Klehr also lists these uses:

• handing repetitive tasks to the Assistant's agent
• asking about facts buried in documents
• summarizing email threads and long video calls, once OpenAI's Whisper has transcribed them
• reminding users to follow up on unanswered emails

The privacy point matters for clients like CSIC, which gives roughly 17,000 researchers a self-hosted space for research data. Data like that may not belong with an outside AI provider.

Klehr says Nextcloud is working to extend its AI features a lot. He wants stronger tool use and multimodal support, meaning models that handle more than text. Ai2 says it is developing models in those areas.

The source gives no performance benchmarks for Olmo 3 against other models. Klehr's hope that more customers would adopt AI depends on a fully open model being equally capable, and the post does not test that. The tool use and multimodal features are plans, not releases.