Practice / Private AI
Models inside your walls. Evidence around every release.
Acyclic deploys open-weight and frontier models in a client’s own cloud account, VPC, or hardware, with controls designed for the data and decisions at hand.
For teams that need the utility of modern models without surrendering the operating boundary.
The practice
A deployment is more than a model endpoint.
The useful work begins with placement: where the model runs, what it can read, which systems can call it, and how its outputs are evaluated. We build those decisions into the deployment.
Our scope can include inference audit trails, model and prompt version pinning, evaluation suites, and data residency requirements. The exact controls are agreed for the client environment and use case.
Client cloud account, VPC, or on-prem hardware, with the deployment boundary documented.
Evaluation suites tied to the intended tasks and failure modes.
Audit trails of inference, appropriate to the workflow and agreed data boundary.
Version pinning for models and relevant system components.
Research from our own systems
Large models on constrained hardware.
Our research includes running 500B+ parameter open-weight mixture-of-experts models on single-workstation hardware using expert streaming and speculative decoding, and coordinating multiple models on shared hardware.
This is Acyclic research, not a client deployment or a performance claim.
See field note draftsA conversation with the people doing the work
Place the model where the work belongs.
Tell us the environment, data boundary, and decision the model needs to support.