Compute you control, running a model you can audit.
Sovereignty asserted in a contract is not the same as sovereignty built into the plant. A SAVRN block is single-tenant, sits behind the customer's own boundary, runs a model whose training data is published, and keeps every action on a ledger a person can read.
Updated
The boundary
A sovereign tenant is physically and logically isolated. The compute is not shared, the network is private, and data is stored in the same facility where it is used. Access is scoped to the tenant's own users and policies, and the tenant decides when and how SAVRN may reach its systems.
Inside the boundary the tenant keeps a private model registry, so the models that run are the models the tenant approved, at the versions the tenant recorded.
- Single tenant per block; no co-location.
- Private network; data stays in the facility where it is used.
- Tenant-scoped access, policies and model registry.
The model
SAVRN is built on Ai2's OLMo 3, a fully open base model with published weights, training data, and post-training pipeline. SAVRN post-trained OLMo 3 on its own domain data using the OLMo 3 model flow (SFT, DPO, RLVR), and operates that post-trained model as the reasoning core of the SAVRN platform. SAVRN is preparing to fork the OLMo 3 lineage into its own model line, taking the SAVRN-tuned model forward under independent naming and release cadence.
That matters for procurement as much as for research. An institution that has certified a model for clinical or controlled work cannot have the base shift underneath it on someone else's schedule. A fork converts that risk into a release cadence the institution can plan around.
- Fully open base: weights, training data and training code published.
- Post-trained by SAVRN on SAVRN data; runs on SAVRN hardware.
- Versioned releases the tenant can revalidate against.
The governance
The AI workforce does the volume: research, drafting, assembly. People hold the gate on anything that ships, signs or spends. Every agent action, approval and document flow writes to a tamper-evident audit trail, and the work itself is metered by the token, so the bill and the ledger describe the same thing.
Work Credits are the unit of value across the platform. A tenant buys capacity, spends it on governed work, and can trace every credit to the work order that consumed it.
- Human approval on every deliverable that leaves the boundary.
- Tamper-evident audit trails for agents, approvals and documents.
- Metered by the token; Work Credits as the unit of value.
Key details
- One tenant per Atom, physically and logically isolated.
- Nothing a customer works on is sent to an outside AI provider.
- A post-trained, fully open base model with an auditable lineage.
- Private model registry and versioned releases inside the boundary.
- Human gates on anything that ships, signs or spends.
- Token metering and a ledger that reconcile to each other.
A university, a hospital system or a public agency can run its most sensitive work on compute it controls, in a facility in its own region, with a model it can audit, without exporting its data to anyone.
How it is delivered
Tenant definition
Users, roles, data classes and the approval gates are written down before capacity is provisioned.
Boundary build
The block, the private network and the model registry are provisioned for that tenant alone.
Model release
The SAVRN model is installed at a recorded version with its evaluation results.
Governed operation
Work orders run through the platform with human approval and full audit.
Revalidation
New model releases arrive on a published cadence the tenant can plan around.
Questions people ask
Is customer data ever sent to an outside AI provider?
No. SAVRN runs its own model on its own hardware inside the tenant's boundary. Nothing a customer works on leaves for an outside provider.
Whose model runs the platform?
SAVRN's own. It is built on Ai2's OLMo 3, a fully open base model, post-trained by SAVRN on SAVRN data, and operated as the reasoning core of the platform. SAVRN is preparing to fork that lineage into its own model line.
Can the tenant bring its own models?
Yes. The private model registry inside the boundary holds the models the tenant approves, at recorded versions.
Who approves what the AI workforce produces?
People do. Every deliverable that ships, signs or spends passes a human gate, and the decision is written to the audit trail with the actor and the evidence.
See Sovereign compute in the whole system.
Every piece here is rendered and narrated on the System page, and the platform that runs it has a guided walkthrough.