You hold the weights, in your own instance, under Australian law
We take the open-weights model that suits your work, train it on your material and run it in Melbourne or your own data centre.
We start from open-weights models released by these labs, so your model doesn’t depend on one lab or one country.
- Mistral AI
- NVIDIA
- OpenAI
- Alibaba
- Moonshot AI
- Z.ai
Your model
A model trained on your organisation’s work.
We align it to Australian law and your policies, and each new version trains on your team’s corrections before you approve it.
Your instance
Each request stays inside your instance.
Sign-in, safety checks, search and the model all run in your instance, whether it’s in Melbourne or your own data centre.
Your law
Stored in Australia isn’t the same
as under Australian law.
A US provider’s Australian region stores your data onshore, but US law can still require that provider to hand it over.
A US provider’s Australian region
- Your data is stored onshoreAustralian data centre
- The provider answers to US lawwherever the data sits
- It can be compelled to hand it overUS CLOUD Act · 18 U.S.C. § 2713
Maincode, in Melbourne
- Your data stays in your instanceMelbourne, or your own data centre
- The provider is Australian-ownedno foreign parent to compel
- Australian law, written into the contractthe law you already answer to
No, I cannot guarantee it.
Who it’s for
For organisations that answer to the public or a regulator.
APRA has warned that a process depending on one offshore AI model can be disrupted by a regulatory decision made overseas.
Government
Keep critical processes on weights you hold and hardware in Australia, where an overseas order can’t suspend them.
Regulated industries
Show your regulator where your AI runs and who can change it.
Enterprises
Turn your people’s knowledge into a model whose weights you own.
Education and research
Keep student and research data in your instance, with test results you can show for each model version.
What you can own
Choose how much of the stack you own.
The options run from Matilda under your own brand to a GPU cluster we build to our reference design and hand over.
-
05
White-label apps
We run Chat, Desktop and the CLI under your brand, on your domain and behind your sign-in.
Your brand and sign-in -
04
Matilda API Where we suggest starting
We run the full API on your hostname, with key issuance, usage metering and audit.
The service -
03
Serving stack
We tune and benchmark serving for your GPUs, and supply quantised variants of the model.
The cluster -
02
Weights
You receive the weights, the tokeniser, the licence to run them and reference configurations.
Everything -
01
Hardware
We design, build and validate a GPU cluster to our reference design, then hand it over.
We build it, you own it
Questions
What to know before you scope a model.
For anything not covered here, talk to our sales team.
Where does our instance run?
It runs in our Melbourne facility or in your own data centre, connected or air-gapped. Compare the options.
What hardware do we need?
You need anything from one multi-GPU node for a team to a full rack for an organisation, and in scoping we check your workload against the hardware you already own. We can also design, build and validate a cluster to the same reference design as MC-1 and MC-2.
Which models can we run?
You run the open-weights model that suits your work, aligned and trained by the Matilda training system at the size your hardware can hold. We size it against your GPUs during scoping, because a size quoted here could be wrong for your rack.
What happens to the weights if we leave?
You keep them, because you hold them under the base model’s licence, and you can run them on any hardware that can hold them.
How do updates work?
Your team’s corrections and confirmations, and any material you add, train the next version, which goes live only when you approve it. Air-gapped sites receive signed artefacts offline, with release notes, evaluation results and a rollback.
Who can switch it off, or restrict our access?
On your own hardware, only your organisation can. In Melbourne we run it under a contract governed by Australian law, and because you hold the weights you can move it. When a US order suspended a frontier lab’s newest models for foreign users on 12 Jun 2026, Matilda kept running, because it doesn’t depend on models hosted in the US.
Do you see any of our traffic?
Not when it runs on your own hardware, because there’s no network route from us to it, which your assessors can verify. In Melbourne, our engineers reach your data only after a paired review, and the access leaves an audit trail. See Trust and security.
How is personal information handled?
Your instance masks personal information on the way in and keeps it out of training. The contract keeps chat history, usage, data and IP in your instance, in Australia, and out of any other customer’s model.
Can our software agents use it, not just our people?
Yes. Agents use the same API as your people’s apps, with tool calling and orchestration included.
What does it cost?
It depends on the level, the model, the hardware and where it runs. For a small pilot a public API costs less, and at scale an owned model usually costs less, so we model both with you before you commit. In our Dec 2025 benchmark, an owned AMD MI355X cluster cut the cost per token by up to 87% against public APIs, for requests of 8,000 input and 1,000 output tokens, with the hardware amortised over three years.
Talk to us