Trellis Data's Agentaus 1: is Australian sovereign AI enough for NZ?

What Trellis Data released

Canberra AI company Trellis Data has launched Agentaus 1, which the company calls a near-frontier sovereign model. Trellis Data's newsroom lists the launch on 10 November 2026, and trade coverage followed on 12 and 13 November. According to Australian Cyber Security Magazine, the model has 220 billion parameters, roughly 50 per cent more than Trellis's previous model, and the company says it runs at scale on smaller compute clusters. Trellis says the model ranks 43rd in the world on the MMLU-Pro benchmark. That figure comes from the company and no independent party has published it.

Agentaus 1 sits inside Agentaus, the platform Trellis took to market in May. In its May launch release, Trellis says model inference, data storage and processing all happen in Australian data centres. The release also says prompts, documents and outputs in Protected Mode aren't used for training, and that on-premises deployment on air-gapped hardware is available. Startup Daily reports that public servants use the platform to turn policy submissions into ministerial briefs that reference the original documents. Access comes in free, subscription, enterprise and API tiers. Trellis hasn't published prices.

The launch coincides with Canberra's plan to legislate mandatory safety and security obligations for frontier AI companies operating in Australia. Technology Decisions quotes founder Michael Gately saying Agentaus 1 and the company's other products "already meet these standards". The legislation hasn't been drafted yet, so nobody can test that claim today.

Why an Australian model matters on this side of the Tasman

Most NZ organisations reach large language models through US companies, and the hyperscaler regions in Auckland don't yet cover the newest models well. AWS's Claude Opus 5 model card shows Auckland (ap-southeast-6) with no In-Region or Geo support. Requests from Auckland can only use the Global profile, which can route anywhere in the world. Sydney and Melbourne can use an Australian Geo profile that keeps processing between those two regions. When AWS opened Bedrock in Auckland, the company described an Australian profile that could keep NZ requests in Auckland, Sydney or Melbourne. Newer models don't always get the same treatment, so check each model separately.

Agentaus belongs to a small group of providers that are Australian-owned and keep both hosting and processing in Australia. Trellis describes the platform as fully Australian-owned and operated. For an NZ firm, an Australian-owned provider takes a US parent company out of the supply chain. Your data still crosses the Tasman, though. A Sydney data centre is offshore under New Zealand law in exactly the way a Virginia data centre is.

What the Privacy Act asks of NZ firms

The Privacy Act 2020 doesn't stop you from using an offshore AI service. According to the Privacy Commissioner's guidance, section 11 treats information that a provider holds or processes on your behalf as information your organisation still holds. You stay responsible for the information, and you usually don't need model contract clauses. The arrangement changes if the provider uses or discloses the information for its own purposes. In that case, both organisations hold the information, and the cross-border rules in Information Privacy Principle 12 can apply. IPP 12 then asks whether the recipient is subject to the NZ Act, protects the information adequately (for example through model clauses), or operates under comparable privacy laws.

The no-training promise in Agentaus's Protected Mode therefore does real legal work for an NZ customer. Whether the promise covers every tier is a contract question. Trellis publishes the commitment for Protected Mode, and NZ buyers should get written confirmation that the free and API tiers work the same way before staff paste in client data.

Australia's privacy law is a reasonable candidate for comparable safeguards. Still, an NZ organisation relying on that argument should write the assessment down rather than assume it. Te Tiriti obligations, sector rules and client contracts often go further than the Act, especially for health, iwi and government data. Many government agencies and health providers have contracts requiring data to stay in New Zealand, and no Australian service can meet that requirement by design.

The on-premises option is the more useful part for NZ

The air-gapped, on-premises deployment option deserves more attention from NZ buyers than the hosted service does. If a sovereign model runs on hardware in your own server room or an Auckland colocation facility, the cross-border question disappears for that workload. Trellis hasn't said whether Agentaus 1 is available on-premises or which hardware it needs. The May release describes on-premises deployment for the platform, and the November coverage only says the new model suits smaller clusters. A 220-billion-parameter model on a small cluster still needs serious GPU memory, so ask for the exact hardware specification before you plan around it.

We see the same trade-off with open-weight models on a local AI server. A smaller model you control can often handle document summarisation, drafting and internal search well enough. Frontier-scale models stay hosted for tasks that need deeper reasoning. Agentaus 1 adds a hosted option between the two, with an Australian owner, and NZ firms can benchmark it against both.

What to do next

• Classify your AI workloads by data type. Separate public material, internal business data, personal information and contract-restricted data. Only the last two groups need a residency decision.
• Read your client contracts. Look for clauses that require data to stay in New Zealand. Those workloads rule out any Australian service, including Agentaus, and need an NZ-hosted or on-premises model.
• Check the inference profile for each model. On Bedrock, look at each model card for Auckland before assuming requests stay in Australasia. Global routing doesn't meet a residency requirement.
• Get the Agentaus terms in writing before a trial. Confirm the no-training commitment for the tier you'll use, where backups live, who can access data remotely, and whether Agentaus 1 runs air-gapped.
• Run your own evaluation. A company-reported benchmark rank won't show how a model handles NZ terminology, te reo Māori or your own documents. Test 30 to 50 real tasks against your current model before switching.
• Record your IPP 12 reasoning. If an offshore provider may use data for its own purposes, write down which IPP 12 ground you rely on and review the record when the provider's terms change.

If you're weighing sovereign, hosted and on-premises AI options together, our AI consultancy team can help you map workloads to the right model and hosting.