What "sovereign AI" really means, and how to test it
What sovereign AI really means in Australia, what Trellis Data's Agentaus 1 claims still need to show, and how I'd test any sovereign model like a POC.

Trellis Data launched Agentaus 1 last week and called it a “near-frontier sovereign AI model” (Trellis Data, 2026e).
My first honest reaction: it depends on the quality of the model. And I didn’t know enough about it to judge.
So I asked for research first. Here’s what I learned, what I can’t verify yet, and how I’d test a “sovereign” model before showing it to a client.
Start with the definition
“Sovereign” gets used as if it were one property. It’s at least three:
- Where your data lives. Is it stored and processed onshore, and does any of it go offshore?
- Who built the model. Was it trained from scratch or adapted from someone else’s base model? On what data?
- Who runs it. Who operates the infrastructure, who are the sub-processors, and who can audit them?
A vendor can be strong on one and silent on another. Residency is the easiest to promise. Provenance is the hardest to check.
The label is slippery. In 2025, Melbourne’s Maincode announced Matilda as Australia’s “first built from scratch, totally sovereign foundational AI model”, then distanced itself from the label by launch (Williams, 2025).
The National AI Plan says locally developed models “may provide benefits such as reflecting unique cultural context or language, supporting innovation and mitigating risks around data security” (Department of Industry, Science and Resources, 2025). Note the “may”.
What Trellis Data says
Trellis Data is a Canberra company, active since 2016 (Australian Business Register, n.d.), whose research team works with the Australian National University (Trellis Data, n.d.-e). It says:
- Size and score. 220 billion parameters (Trellis Data, 2026e), and 84.4% on the MMLU-Pro benchmark, which “at the time of testing placed it 43rd globally”, in results Trellis will look to publish soon (David, 2026).
- Its own model. Region Canberra reports that the first Agentaus, in May, adapted open-source frameworks, but Agentaus 1 is Trellis’s “proprietary 220-billion-parameter core model” (David, 2026). Founder Michael Gately put it plainly on X: “it’s our own proprietary model, trained in Australia” (Gately, 2026a).
- Data stays home. Agentaus runs on Trellis’s own cloud in Australian data centres, “not AWS, Azure, or other public hyperscalers” (Trellis Data, n.d.-c). Customer data is never used “to train, improve or fine-tune our underlying AI models” (Trellis Data, 2026a).
It’s also easy to try, with a free tier and an API Trellis says is compatible with OpenAI’s chat completions API (Trellis Data, n.d.-a, n.d.-d).
What I couldn’t verify yet
None of this means the claims are wrong. Some just can’t be checked from outside yet.
The model itself. There are no published weights, model card or technical report. In June, a reviewer noted that Trellis “will not say which underlying model powers the assistant” (Adam, 2026). The Agentaus FAQ, still live this week, says Protected Mode uses “customised open-source models hosted by Trellis Data” and that Trellis “is also building its own frontier model” (Trellis Data, n.d.-c). That page may predate the launch. Worth asking.
The benchmark. Agentaus 1 doesn’t appear in the public data behind TIGER-Lab’s MMLU-Pro leaderboard, last updated in March 2026 (TIGER-Lab, 2026). That fits Trellis saying results are still to come. Until then, it’s self-reported.
A coincidence, nothing more. 84.4% is exactly the MMLU-Pro score Alibaba’s Qwen team reports for its open-weight Qwen3-235B-A22B-Thinking-2507 model (Qwen Team, n.d.).
What does that mean? Nothing, yet. Different models can land on the same number on one benchmark. Trellis says its model is its own and trained in Australia.
It isn’t proof of anything, including that Agentaus 1 is built on Qwen. The useful response is the same as for any self-reported score: ask for the evaluation settings and an independent check.
“Nothing ever leaves Australia.” That’s how Gately put it on X this week (Gately, 2026b). When Agentaus first launched in May, Trellis tied its zero-training promise to “our Protected Mode” (Trellis Data, 2026c). Today’s plans, including the free one, also list an “External Mode” with “global engines, results kept protected” and a monthly top-up of “Claude & ChatGPT credit” (Trellis Data, n.d.-a).
I’m not saying your data leaves the country. I’m saying you should know which mode your people are in, and whether you can switch External Mode off.
The operator. Trellis doesn’t disclose sub-processors “for security reasons”. Its Essential Eight Maturity Level 2 is self-assessed, with an independent assessment scheduled, and IRAP-assessed infrastructure is “available on request” (Trellis Data, n.d.-c). Reasonable starting points, not audit reports.
Who builds it, and what goes in?
This worries me most, and it isn’t specific to Trellis. Who is creating the model? Are they injecting their own biases into it?
Trellis hasn’t published what Agentaus 1 was trained on. That’s common in this industry, but it means you can’t read the model’s values off a datasheet.
There’s also a layer you can’t see. The developer docs say your system message “is layered on top of the built-in system prompt”, and “there is no way to fully override it” (Trellis Data, n.d.-b). In the docs’ own example, a request that says only “Hello” counts 2,440 input tokens.
That isn’t sinister in itself, but it shapes every answer. And if a model is adapted from another base, it can carry some of that base’s habits, wherever the base comes from. So probe sensitive political topics from several countries, not just one.
The bubble question
My instinct was that a model created primarily for the Australian market, with our laws and regulations, might be blind to what’s happening outside Australia. A bubble solution: it works inside the bubble, but ask about something else and it may not have the information. Like a model that only speaks English.
Honestly, the research suggests the opposite risk may be bigger. If the 84.4% holds up, the model has broad general knowledge, because MMLU-Pro covers about 12,000 questions across 14 subject areas, most of them nothing to do with Australia (Wang et al., 2024). Trellis also says it works “across English, Indonesian, Chinese, Arabic and more” and can search the internet (Trellis Data, n.d.-a, n.d.-c). The real gap could be thin Australian detail or an old knowledge cut-off.
Either way, it’s testable. And even within Australia we have nuances of culture, language and ideas.
Test it like any software POC
This isn’t different from running a proof of concept for any software. You try it with real data, or close to real data, and see how it reacts to the sovereign questions and Australian law.
One catch: my first thought was to install it and review it. You can’t. There are no downloadable weights, and the user agreement bars copying, derivatives and reverse engineering, other than as permitted by law (Trellis Data, 2026b). So you test through the app and the API, in this order:
- Legal first. Review the user agreement and privacy policy before anyone signs up with a work identity. The OAIC recommends not entering personal information, particularly sensitive information, into publicly available generative AI tools (Office of the Australian Information Commissioner [OAIC], 2024).
- Ask in writing. Can External Mode and web search be turned off for your whole organisation? Who are the sub-processors? Is there an IRAP report for the service, and what’s its scope? What’s the base model?
- Start near-real. Use synthetic or de-identified documents, and move to real data only once steps 1 and 2 are cleared.
- Run the same tasks through two other models, for example an onshore hyperscaler model and an open-weight model you host yourself.
- Score with people, on your work. Benchmarks don’t transfer neatly to internal tasks. Even MMLU-Pro scores move about 2% with prompt style (Wang et al., 2024).
- Probe the bubble and the bias. Pair questions on Australian and non-Australian law and current affairs, and use dated events to find the knowledge cut-off.
Regulated buyers have homework. Where a third party manages your information assets, APRA’s CPS 234 says you “must assess the information security capability of that party” (Australian Prudential Regulation Authority [APRA], n.d.). A web page isn’t that.
Why I’d still run more than one model
Trellis pitches Agentaus 1 as “reducing the need to access multiple AI models for different tasks” (Trellis Data, 2026e). I see it differently.
I think the future is multi-model. Companies don’t run one piece of software for everything. They take a multi-software or multi-cloud approach, because they don’t want to put all their eggs in one basket. Using different vendors reduces the risk, and AI is very similar. Databricks’ AI Index, drawn from more than 10,000 companies using its Unity Gateway, found the share relying on a single model provider fell from about 64% in July to 28% by early November 2026, and about 71% had used two or more providers (Databricks, 2026a, 2026b). That’s one vendor’s customers, not all enterprises, but it points the same way.
It helps with bias, too. You can balance one model that might have biases or opinions you don’t know about against another that’s more open on that topic. And it gives companies time to reflect, review, evaluate and decide.
The federal GovAI platform is designed to be “vendor-agnostic”, with a model selection option based on need (Department of Finance, n.d.). Trellis’s own partner announcement says Workato is “model-neutral, so organisations can run sovereign and frontier models side by side” (Trellis Data, 2026d).
I’ve argued that model choice is turning into a routing decision, not a vendor decision. A sovereign model can be a good route for sensitive workloads. That doesn’t make it the only route.
What I’d tell another practitioner
- Ask which layer of sovereignty you’re buying. Residency, provenance and operational control are different promises.
- Treat self-reported benchmarks as a starting point. Wait for independent results, or run part of the public MMLU-Pro benchmark yourself (Wang et al., 2024).
- Ask for artefacts, not adjectives. An IRAP report with its scope, an independent Essential Eight assessment, a sub-processor list, a model card.
- Keep at least two models in play. An OpenAI-compatible API should make switching easier, so design for it from day one.
So, is it sovereign?
On data residency, Trellis Data makes clear, specific promises, with one mode that needs a closer look. On provenance, we don’t know yet. On operations, it’s a sensible start, with artefacts still to come.
I’d welcome an Australian model that holds up. But before I trust the claims, I need independent benchmarks, independent reviews, maybe some papers. More people will be looking at this, and I’ll keep watching.
Would you trust a “sovereign” label before you’d tested what’s under it?
~ Mr. Gómez
References
Adam. (2026, June 23). Agentaus review (2026): Australia’s sovereign AI assistant tested. The AI Directory. https://www.theaidirectory.com.au/review-agentaus-australia-2026/
Australian Business Register. (n.d.). ABN 67 616 516 619: Trellis Data Pty Ltd. Retrieved November 18, 2026, from https://abr.business.gov.au/ABN/View?abn=67616516619
Australian Prudential Regulation Authority. (n.d.). Prudential Standard CPS 234 Information security. Retrieved November 18, 2026, from https://www.apra.gov.au/standards/cps-234
Databricks. (2026a). Databricks AI Index [Interactive dashboard]. Retrieved November 18, 2026, from https://developers.databricks.com/ai-index
Databricks. (2026b, November 16). Introducing the Databricks AI Index: The state of enterprise AI tokenomics [Blog post]. https://www.databricks.com/blog/introducing-databricks-ai-index-state-enterprise-ai-tokenomics
David, D. (2026, November 16). Australia’s own AI powers up to rival global giants. Region Canberra. https://region.com.au/australias-own-ai-powers-up-to-rival-global-giants/1014588/
Department of Finance. (n.d.). About GovAI. Retrieved November 18, 2026, from https://www.govai.gov.au/about
Department of Industry, Science and Resources. (2025). National AI plan. Australian Government. https://www.industry.gov.au/sites/default/files/2025-12/national-ai-plan.pdf
Gately, M. [@MGately_au]. (2026a, November 18). Thanks @chrissamanyy! Small note: it’s our own proprietary model, trained in Australia [Post]. X. https://x.com/MGately_au/status/2122820166446891379
Gately, M. [@MGately_au]. (2026b, November 18). We built Australia’s own AI model, and today it’s open to the world [Post]. X. https://x.com/MGately_au/status/2122780345947017287
Office of the Australian Information Commissioner. (2024, October 21). Guidance on privacy and the use of commercially available AI products. https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products
Qwen Team. (n.d.). Qwen3-235B-A22B-Thinking-2507 [Model card]. Hugging Face. Retrieved November 18, 2026, from https://huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507
TIGER-Lab. (2026). MMLU-Pro leaderboard submission results [Data set]. Hugging Face. https://huggingface.co/datasets/TIGER-Lab/mmlu_pro_leaderboard_submission
Trellis Data. (n.d.-a). Agentaus. Retrieved November 18, 2026, from https://trellisdata.com/agentaus/
Trellis Data. (n.d.-b). Chat completions [Developer documentation]. Agentaus. Retrieved November 18, 2026, from https://agentaus.com.au/eng/docs/developers/chat-completions-api
Trellis Data. (n.d.-c). Frequently asked questions. Agentaus. Retrieved November 18, 2026, from https://agentaus.com.au/eng/docs/faq
Trellis Data. (n.d.-d). Getting started [Developer documentation]. Agentaus. Retrieved November 18, 2026, from https://agentaus.com.au/eng/docs/developers/getting-started
Trellis Data. (n.d.-e). Research. Retrieved November 18, 2026, from https://trellisdata.com/research/
Trellis Data. (2026a). Privacy policy (effective April 28, 2026). https://agentaus.com.au/eng/docs/privacy
Trellis Data. (2026b). User agreement / licence agreement (effective April 28, 2026). https://agentaus.com.au/eng/docs/agreement
Trellis Data. (2026c, May 7). Launch of Agentaus: Australia’s sovereign, high-security AI platform. https://trellisdata.com/newsroom/trellis-data-launches-agentaus/
Trellis Data. (2026d, October 14). Trellis Data’s Agentaus powers sovereign AI model Dukes, now integrating with Workato. https://trellisdata.com/newsroom/agentaus-powers-dukes/
Trellis Data. (2026e, November 10). Trellis Data joins global AI ranks with launch of near-frontier sovereign AI model: Agentaus 1. https://trellisdata.com/newsroom/trellis-data-launches-agentaus-1/
Wang, Y., Ma, X., Zhang, G., Ni, Y., Chandra, A., Guo, S., Ren, W., Arulraj, A., He, X., Jiang, Z., Li, T., Ku, M., Wang, K., Zhuang, A., Fan, R., Yue, X., & Chen, W. (2024). MMLU-Pro: A more robust and challenging multi-task language understanding benchmark (arXiv:2406.01574). arXiv. https://doi.org/10.48550/arXiv.2406.01574
Williams, T. (2025, October 15). Maincode debuts Matilda AI after ditching “sovereign” label. Information Age. https://ia.acs.org.au/article/2025/maincode-debuts-matilda-ai-after-ditching—sovereign—label.html