Frequently Asked Questions

A grassroots collective of European researchers, engineers and volunteers building open-weight AI models. We coordinate talent and pool compute (both from volunteers and partner institutions) to ship models, datasets and evaluations that anyone can use.

No. Right now it's an informal collective. When we need to hold funds or sign contracts, we'll set up a lightweight legal entity (likely a Dutch stichting or Belgian AISBL). But we start lean: structure follows shipped work.

The product isn't the chat, it's the shipped artifacts: models, training recipes, evaluations, datasets. We use Discord (or Matrix) for coordination, but what matters is what we release publicly.

Open Research Collective for AI. And yes, it contains all five vowels. Easy to remember, easy to say.

Yes, for many workloads. Distributed training frameworks like DiLoCo (from DeepMind), Hivemind, and Petals enable training across machines connected over the internet, not just in a single data center.

Not everything works on volunteer hardware. Frontier pretraining of the largest models still needs concentrated compute. That's why we have two lanes: volunteers handle evals, data processing, distillation, and smaller training runs. Partner clusters handle the heavy stuff.

Safety is a core design requirement, not an afterthought:

• Signed workloads: your machine only runs verified, auditable job bundles
• Sandboxed execution: workloads run inside containers with additional isolation (gVisor/Kata)
• No secrets on your machine: volunteer tasks never require credentials or sensitive data
• Secure updates: the client uses signed releases with a clear revocation path

We assume some nodes will be malicious. That's the design baseline. We use redundant execution (same task on multiple machines), spot-checking, quorum rules, and deterministic pipelines where possible. BOINC and Folding@home have validated these patterns for 20+ years.

If we can't train our own models, we depend on US and Chinese companies for:

• Core AI capabilities and access policies
• Safety standards and alignment research
• Multilingual and European-language support
• Bargaining power on compute infrastructure

This isn't nationalism, it's resilience. Open-weight models that Europe can train, inspect and deploy independently reduce structural dependency.

Yes, and that's great. OpenEuroLLM (launched 2025) is a 20-institution consortium building open European LLMs. EuroHPC is deploying AI Factories across the continent. The EU AI Act creates a regulatory framework.

But institutional programmes take time (12–18 months to funding), have high rejection rates (~85%), and come with significant admin overhead. ORCAI complements these efforts: we move faster, with lower overhead, and keep everyone honest through open publication.

No. We need engineers, data specialists, community builders, writers, and anyone willing to donate GPU cycles. Work matters more than credentials.

We're Europe-focused because that's where the coordination gap is biggest, but anyone aligned with the mission is welcome. Open means open.

Every contribution is logged publicly. We use CRediT-style roles (the same taxonomy used in academic publishing) so your work is visible and attributable. No honorary authorship. If you did the work, you get the credit.

Not in v1. ORCAI starts volunteer-first to keep agility. Down the road, we may offer micro-grants (€500–5,000) for specific needs like compute, data acquisition or travel. But the first currency is impact and recognition.