Mentary AI: Verifying a Private Inference Deployment on a Fresh Host

Description and use

Verifying Mentary AI local deployment: network isolation test, package inventory, health and load checks, compliance assessment, and price validation.

Problem

Verify that a purchased Mentary AI package truly runs locally without external API calls and contains all promised components before deployment.

Outcome

The deployment runs locally without external calls, all promised artifacts are present, performance meets expected latency for single concurrent user, and compliance benefits are confirmed.

Workflow

• Run network isolation test using unshare -n to ensure no outbound traffic during inference.
• Inventory package contents with du -sh /opt/mentary/weights /opt/mentary/runtime and list deployment scripts and Docker images.
• Perform health check with curl http://127.0.0.1:8080/healthz and load check measuring latency with curl -w '\n%{time_total}\n'.
• Review compliance claims: confirm absence of third‑party processors, request BAA/SOC‑2 if needed during support window.
• Verify pricing and payment options as listed on the product page.

Evidence

Block all outbound traffic for this process, then run an inference.

Source

If a response comes back, the model is genuinely local. If it hangs or errors on DNS, you have a hidden call out — which is the single thing that would disqualify the product for a privacy buyer.

Source

Before any tuning, inventory it. The page lists: model weights, the optimized inference engine, deployment scripts, a production Docker image with Compose orchestration, documentation, and 30-day priority support.

Source

The page says CPU and GPU-accelerated runtimes are both included. That means two different performance profiles, and you should measure the one you will actually run.

Source

What you do not get is a BAA, a SOC 2 report, or a signed residency guarantee.

Source