Free tools · sovereign AI
Plan AI you can keep control of.
Two private checks that run only in your browser. One maps where AI data and model artifacts travel; the other estimates whether a model fits the memory you have. Treat them as planning aids. They aren’t security audits or deployment guarantees.
01 / Readiness check
Sovereign AI exposure checker
Answer based on the system as it runs today. The result highlights unanswered control questions; it is not a legal or compliance certification.
Choose the closest answer in each field to see control gaps to investigate.
A higher score only means more answers describe documented organizational control. It does not establish data residency, sovereignty, security, or compliance. Validate contracts, subprocessors, backups, telemetry, identity, and applicable rules with your security and legal teams.
02 / Model fit
Local and on-prem model memory estimator
Estimate weight memory for a dense model from parameter count and quantization, then add context-cache and runtime headroom. Actual use depends on architecture, runtime, context, batching and offload behavior.
Adjust the assumptions and estimate. This is not a performance benchmark.
Planning heuristic: parameter count × bits ÷ 8 estimates packed weight size, with 20% added for quantization metadata. The rough KV-cache allowance is 0.5 GB for an 8-billion-parameter model at 8,192 context tokens, scaled linearly with parameter count and context length, then 20% operating headroom is added. This is not a model-specific KV-cache calculation. MoE models, multimodal encoders, adapters, batching and runtime-specific allocations can differ materially. Treat a near fit as uncertain and benchmark the exact model and serving stack.
Primary references (checked 2026-10-23)Model size and license claims change; follow the linked model card and license for the exact revision you download. This page makes no license or hardware compatibility determination.
Need to discuss a controlled deployment? Talk to the Hyperlake team.