Mentary AI on Two Hosts: What I Actually Had to Do to Get a Failover Path Working

Mentary AI at $299.99 is a one-time license for a fully private, autonomous local AI model: model weights, inference engine, deployment scripts, a Docker image, docs, and 30 days of priority support. That part is well covered.

What almost nobody plans for is the second host. Below is what I actually had to sort out when I put a second box next to the first one.

What "redundancy" means when there is no cloud to fail over to

With a cloud vendor, failover is someone else's problem. With a local model, the failure modes are all yours:

• Host dies. The model was on a disk you own. If that disk is the only copy, you are restoring from backup, not failing over.
• Model file is the single copy. Weights plus runtime on one box. Two hosts that both point at one NFS export are not redundant.
• Serving layer. Something has to accept requests when the primary is gone. The package gives you the model and the runtime; the routing in front of it is the part you build.
• Session state. If your team keeps a working conversation open, a failover that starts a fresh session mid-task is a disruption even though it "worked".

The concrete setup

Two bare-metal hosts, the same model revision loaded on both.

# on each host, before you point anything at it
md5sum ./weights/* | sort          # both hosts must match byte-for-byte
./bin/mentary-health --json      # runtime up, weights mounted, GPU visible
curl -s localhost:8080/healthz      # serving layer answering

Two things worth checking that people skip:

1. Model revision equality. If the second host was set up three weeks later and you re-downloaded, you may be running a different revision than the first. Verify the weights, not just the directory listing.
2. GPU memory headroom. The model loads into VRAM. If host two also runs other work, the model may fit at first and OOM under load.

Where the $299.99 package ends

It covers: model weights and runtime, deployment scripts for bare-metal and VMs, the Docker image with Compose, and documentation including security hardening and performance tuning. 30 days of priority email support for environment-specific issues.

It does not include: the routing layer, shared storage, or the failover policy. You either write that yourself or you bring the deployment engineering support for it.

That distinction matters when you are comparing a $299.99 local license against a cloud subscription. The subscription includes the routing. The local license includes the model and the right to run it anywhere.

Before you buy

• Do you actually need two hosts, or one host plus a tested restore procedure? For many teams, restore is enough and costs far less.
• Do you have a place to put the second copy of the weights that is not the same physical machine?
• Who answers the phone when the primary dies at 03:00? If the answer is "whoever is on call", the failover path needs to be boring enough that anyone can run it.

Try it first

The live demo on the product page lets you run queries locally in your browser before paying anything — you can see the model answer and check that the local-processing claim holds. Worth doing before you commit to hardware for two hosts.

Mentary AI — $299.99, one-time, lifetime license. Model weights, optimized inference engine, deployment scripts, production Docker image, documentation, 30-day priority support. No recurring fees.

Payment: PayPal to [e-mail] (buyer writes a note, seller confirms in the inbox), crypto to 0x078fE079A9440631Bc6AD722D7143dC382b4ea4f on Base (ETH / USDC / USDT), or Western Union / MoneyGram. After paying, email [e-mail] with the transaction hash or MTCN for verification.

Not a capacity planner. These are the failure modes and checks from running two local hosts — your hardware and network will differ, so validate on yours.