hangar. Get early access
Private preview

Your AI application.
Running in Australia.

Bring your model or choose one from our catalogue.
See which Australian runtime meets your quality and speed targets at the lowest measured cost. Get a signed record of the run.

Get early access Explore the platform

Tested in Sydney and Melbourne by TALOS Defence.

Meet the company
Compute cost

About two-thirds lower

Compute cost per correct call summary in our synthetic banking test, comparing our initial and tuned Trainium2 setups in Melbourne.

Output agreement

187 of 192 matched exactly

In a separate test, the Sydney H100 and Melbourne Trainium2 produced byte-identical summaries for the same synthetic calls at eight concurrent requests.

Run evidence

Verifiable offline

Download the signed record. A reviewer can check the recorded release and environment even after the compute has shut down.

What counted as a correct summary?

In our 25 October 2026 test, a synthetic call had a known answer for each required field. A wrap-up had to return every field correctly to count. We compared compute cost per 1,000 accepted wrap-ups across configurations that met the same five-second p95 response-time target. The separate cross-city figure measures agreement between outputs. Customer pricing and savings depend on traffic and the checks the workload must pass.

Request the engineering evidence

What will it cost to run your app here?

A GPU hourly rate only gets you so far. Hangar measures how much useful work you get for that hour, at the speed your users need.

Choose where it runs.

If a customer requires Melbourne, put that in the brief. Review the locations for compute, storage, logs and backups before the workload runs.

Data location

Cut the cost per useful answer.

Give us examples of a good result and a response-time limit. Our specialist engineers compare supported runtimes at your expected load and show you what each passing output costs.

Inference economics

Give your reviewer a receipt.

When a customer asks which release ran and where, hand them the signed record. They can verify it on their own machine without access to your cloud account.

Independent verification

Pick a catalogue model.
Or bring your own.

Tell us what goes in and what a good result looks like. We’ll confirm which model and city support it in the private preview.

Reasoning and chat
Answer questions about a case file. Test the answers against examples your subject-matter experts have approved.
Speech and call summaries
Start with audio for transcription or an existing transcript for wrap-up. Check whether the summary captures the issue, outcome and follow-up.
Documents and images
Extract the fields your application needs from forms, scans and correspondence. Check the values against a known answer.
Classification and triage
Return a category your software can act on, such as “needs review”. Measure missed cases alongside the cost per decision.

Bring one workload.

A description is enough for the first conversation. Keep customer data and model files in your environment until we agree how the trial will handle them.

Tell us what you’re running
  1. 01

    Set the acceptance test

    Choose a representative task, expected traffic, response-time target and the region your customer needs.

  2. 02

    Compare it with today’s setup

    Review cost per accepted result, latency and examples that failed. See whether switching pays at your volume.

  3. 03

    Choose the runtime and keep the evidence

    You approve the environment for the preview run. Your team gets the signed record to verify and share with its reviewers.

Before you bring a workload

How do I get started?

Request an invitation with your email. When we follow up, a description of the task, its volume and any location requirements is enough to check fit. Keep customer records and model weights out of the enquiry.

Can I bring a model I already use?

Yes. Bring an open-weight model or a fine-tune you have the rights to run. We check compatibility and use examples you approve to compare its outputs with your current application. You can also start with a supported catalogue model. The team confirms model and regional availability during the private preview.

Can I specify Melbourne or Sydney?

Tell us which city your workload requires. Before a preview run we document where inference, storage, logs and backups will sit, who can access them, retention periods and any external tool destinations. You can review those details before providing workload data. The public website signup is covered by its own privacy notice.

How do you work out whether I would save money?

Take a call summary as an example. We count summaries that pass your checks, measure response time at your expected traffic and divide the compute bill by the accepted count. We compare that figure with your current setup and identify standby, storage and other charges separately. That gives you a basis for a buying decision.

What can my reviewer verify?

A signed run record identifies the recorded workload release and execution environment. Your reviewer can check the signature and fingerprints offline and detect changes to the record. Model quality is assessed against the workload’s acceptance tests. Your organisation determines how that evidence supports its obligations.

Who is behind Hangar?

Hangar is a product of TALOS Defence, an Australian-owned and operated technology company working on national security and critical systems. Engineering evidence and further security information are available on request.

Access by invitation

Request early access

Leave your email. We’ll get in touch about the workload you want to run and the Australian region you need.