Frequently asked questions¶
General¶
What is the difference between AIGr.id and AIOS (OpenOS.AI)?
AIGr.id is the network: the clusters, models and people that take part. AIOS, also called OpenOS.AI, is the open-source software that runs it. You can use AIOS to start your own network, private or public, or to join an existing one.
Is it open source? Under what license?
Yes. AIOS is 100% open source and community-driven, released under the Apache License 2.0.
Is it ready for production?
Not yet. A variant of the system has run in production at close to 500,000 inferences per second for about a year. The current release adds major new capabilities, such as LLM support, and has not yet been validated at that scale. It is in beta and not recommended for production use. See the roadmap on the home page.
Who owns the network?
No single party. Each network is governed by whoever runs its management cluster, each cluster by its owner, and each block by its creator. Each of them sets rules through policies. Many networks can exist independently.
For businesses and organizations¶
Can we run a private AI grid just for our organization?
Yes. You can create your own network on your own hardware and control who joins through governance policies. This is the "enterprise sovereign AI grid" use case.
Do we keep control over our data and models?
You decide where your blocks run, through allocation policies; who may use them, through quota and access policies; and which clusters and nodes may join, through pre-check policies. Your models and data stay on the infrastructure you choose.
Do we have to replace our existing AI stack?
No. Existing systems, such as a vLLM server, can be wrapped as blocks. See Third-party systems.
What security features are available?
Platform-wide identity and access management (decentralized identity, role-based access control) and model/asset security are on the roadmap and not yet released. Today you can control actions through pre-check policies and quotas, and you should protect cluster access with a VPN and firewall rules. See Security considerations.
For infrastructure owners¶
What do I need to join?
- To contribute a cluster: a working Kubernetes cluster set up with
kubeadm,kubectland Helm, and the network ports listed in Join a cluster. - To contribute a single machine:
kubeadmat the same version as the target cluster, plus network access. See Join a node.
Which Kubernetes version is required?
AIOS runs only on Kubernetes but does not require a specific version, as long as the cluster is API-compliant. A node joining a cluster must use the same kubeadm version as that cluster.
Do I need GPUs?
No. Nodes with and without GPUs can join. Some blocks need GPUs, and allocation policies place them only where suitable GPUs exist.
Can several workloads share a GPU?
Yes. Several block instances can be placed on the same GPU, if the resource allocation policy allows it.
For developers¶
What languages can I use?
Block code, policies and init containers are written in Python, using the AIOS SDKs. Through third-party integration, a block can call a system written in any language.
Can one block span several clusters?
No. A block can span several nodes within one cluster. To use several clusters, create several blocks and connect them with a vDAG.
My model doesn't fit on one machine. What can I do?
You can split it into parts, deploy each part as a block, and connect them with a vDAG. If the model fits on one machine with several GPUs, a single block can use several GPUs (see the Mistral 7B example in the Quickstart).
How do clients call a block or workflow?
Over gRPC: through an inference server for single blocks, or through a vDAG controller for workflows. See the Inference user guide.
Getting help¶
Where can I ask questions or contribute?
- Discord: OpenCyberspace
- Email: community@opencyberspace.org
- X (Twitter): @opencyberspace