Trellis Data launches Agentaus 1, a 220B-parameter model for sovereign AI

Australia has a new claimant in the race to build AI at home. Canberra-based Trellis Data has launched Agentaus 1, a 220-billion-parameter model it describes as near-frontier and built for “sovereign” use — Australian businesses and government agencies that want their data to stay on Australian soil. The launch lands the same week Canberra flagged new mandatory AI safety and security protections, and Trellis is pitching Agentaus 1 as a model that already meets them. The catch is verification: the headline benchmark claim is the company’s own, and the model’s inner workings remain undisclosed.

What actually shipped#

Agentaus 1 is a 220-billion-parameter model that Trellis says is about 50% larger than its previous model, with claimed improvements in accuracy, multi-step reasoning and output length — and the ability to run at scale on smaller compute clusters. It is available now through the Agentaus platform, in free, subscription, enterprise and API tiers, according to the company. Trellis announced it in its newsroom on November 10. The company also cites a “world ranking of 43” on MMLU-Pro, a widely used benchmark of knowledge and reasoning, as evidence that the model sits near the frontier.

The sovereignty pitch#

The pitch is less about raw capability than about control. Trellis says Agentaus 1 runs with local data hosting and “defence-grade encryption,” backed by a Canberra-based team, and that customer data is never used to train its models. The company describes its background as work in high-security environments across government, intelligence, defence and law enforcement. Founder Michael Gately framed it as proof Australia can produce advanced AI rather than only consume technology built elsewhere, keeping data and investment at home — a matter, he argued, of productivity, prosperity and national resilience.

The timing is deliberate. The federal government flagged plans for new mandatory AI safety and security protections earlier this week, and Trellis says Agentaus 1 and its other products already meet those standards — a claim that turns regulation from a compliance cost into a selling point, if it holds up in procurement. Agentaus is not starting from zero: the platform launched commercially in May 2026, and it already powers Dukes, a sovereign model from Indigenous ICT provider Gulanga Group being integrated with Workato’s orchestration platform.

What Trellis hasn’t shown#

What the launch omits is as important as what it includes. Trellis has published no architecture details, no model card, no training provenance and no independent benchmark results — so the MMLU-Pro ranking, the accuracy claims and the reasoning gains are all the company’s word. An independent review of the Agentaus platform in June 2026 found that Trellis would not disclose which underlying model powers it, describing its architecture only as proprietary, and the reviewer could not verify the base model against any primary source. That opacity matters more for a sovereignty claim than for an ordinary launch: buyers are being asked to trust not just the model’s quality but its provenance.

It is also worth keeping the claim in proportion. A 220-billion-parameter model signals scale, but scale is not proof of trust, resilience or suitability for government work. Parameter count says nothing about how the model was trained, where it runs, who administers it, or whether customer data stays inside a defined boundary. In Australian policy debates, sovereignty is treated as a systems question — infrastructure, governance, skills and operational control — not a single setting. A model can be hosted in-country yet depend on external update channels or opaque administration, while a more isolated deployment shifts patching and monitoring onto the customer. The right question for any buyer is a data-flow one: where do prompts go, where are retrieved documents stored, who can see the logs, and how are updates delivered? Until Trellis answers those in public, the launch is best read as a claim that needs evidence, not as proof.

Why it matters#

Even with the caveats, the launch is a sign of the times. Sovereign AI has moved from conference slogan to procurement criterion: governments are writing mandatory protections, and agencies are asking where their data lives before they ask how smart the model is. Until now, an Australian buyer wanting a frontier-class model had to choose between American and Chinese systems; a domestic 220-billion-parameter option, if it performs as claimed, gives them a third answer. The pattern is global — this week alone, British chip startup Fractile raised $720 million at a $7 billion valuation with UK state backing — and countries that can field credible domestic AI gain leverage over the terms on which they use everyone else’s.

What to watch#

Four questions decide whether Agentaus 1 is a milestone or a press release. First, whether independent benchmarks appear — a public evaluation would settle the MMLU-Pro claim quickly. Second, whether the procurement angle materialises: agency pilots or contracts would be the real test of the pitch. Third, whether Trellis discloses architecture and training provenance, which is what would let buyers actually audit the claim rather than take it on trust. And fourth, pricing: no paid pricing has been published, so the cost case is still unproven. Until those answers arrive, Agentaus 1 is a genuinely interesting launch with an evidence gap: a 220-billion-parameter bet that Australia can be an AI producer, waiting for the receipts.

Sources: Australian Cyber Security Magazine — Trellis Data launches near-frontier model Agentaus 1 (Nov. 12, 2026); Netcrook — Australia’s AI Sovereignty Test Starts With the Plumbing, Not the Hype (Nov. 12, 2026); Trellis Data — Newsroom. Capability and benchmark claims are Trellis Data’s own and have not been independently verified.