An air-gapped LLM, with no internet connection at all.
For the most sensitive workloads. The model runs on a network with no internet access, and every model and update arrives as a signed bundle your team checks before it is installed.
What you get
Everything it needs, inside the gap.
Offline weights delivery
Model weights, container images and dependencies packaged as one bundle, so nothing is downloaded at install time.
Signed bundles and SBOM
Every bundle is signed and comes with a software bill of materials. Your team verifies both before install.
On-prem GPU serving
The serving stack runs on your GPUs inside the isolated network, with a standard API for your applications.
Local evaluation harness
Your test suite runs inside the gap, so every update is tested where it will run.
Offline monitoring
Usage, errors and performance recorded locally, for your team to review.
Runbooks for your team
Install, update, rollback and recovery, written for the people who will do them.
How updates work
Every change crosses the gap the same way.
01
Build
We build the bundle outside the gap: weights, images, dependencies and test results.
02
Sign
The bundle is signed, with a software bill of materials listing what is in it.
03
Transfer
It moves across through your approved media and process.
04
Verify
Your team checks the signatures and the bill of materials before anything is installed.
05
Install and test
Installed by your team, then tested with the local evaluation harness. Rollback is the same process in reverse.
Technical detail
What we install, and how we size it.
Supply chain
Signed container images and bundles, a software bill of materials for each release, and a local registry inside the gap.
signed images · SBOM · offline registry
Serving
vLLM or SGLang for most workloads, TensorRT-LLM where peak throughput on NVIDIA hardware matters. A standard API in front, so your applications don’t change if the model does.
vLLM · SGLang · TensorRT-LLM
Sizing
A 70B-parameter model needs about 140 GB for weights at 16-bit, 70 GB at FP8 and 35–40 GB at 4-bit, plus memory for each conversation in flight. We size from measured throughput at your expected load, and re-run your tests after any quantisation.
FP8 · AWQ · GPTQ
Who has access
No remote access, ever.
An air-gapped system can’t be reached from outside, including by us. Work happens on site under your supervision, or your team does it using our runbooks.
Questions
What security teams ask.
How do updates reach an air-gapped site?
As signed bundles with a software bill of materials. Your team checks the signatures and installs them.
Can you support it remotely?
No. The deployment is designed so nothing inside the gap is reachable from outside, and isolation is verified at install and after each update. Support happens on site, or through your team with our runbooks.
How do you test a model you can’t reach?
The evaluation harness runs inside the gap. Your team runs it after every install and shares the results.
Can we roll back an update?
Yes. The previous bundle stays on site, and rollback follows the same verified process.
What we don’t promise
Some claims we won’t make.
- Fast updates. Each one moves at the speed of your transfer and approval process.
- Remote fixes. If something breaks, the fix is on site or through your team.
Nothing in, nothing out.
We’ll design the update process with your security team before anything is built.