Trellis Data unveils Agentaus 1, a 220B-parameter Australian model pitched as cheaper and locally controlled

Trellis Data, a Canberra-based AI company, will unveil its new Agentaus 1 model on Thursday. The timing is pointed: Anthropic is in Australia trying to convince investors that its AI platforms justify a $2 trillion valuation, while Trellis is offering what a local firm with under $10 million in external funding can build.

By size, Agentaus 1 sits at around 220 billion parameters, about 50 percent larger than Trellis's previous model. That is far smaller than the frontier systems from US labs, and founder Michael Gately does not dispute it. He frames the gap in time instead: Trellis's models are less than a year behind Anthropic's in capability.

His self-assessment is blunt. “On benchmarking we are sitting at 43rd in the world, which may not seem super high, but we are lagging the frontier models by four to six months and customers are getting just as good results with us for what they need to do,” he said. That sentence is essentially the business case: not the strongest model, but good enough across most professional tasks.

Trellis is not a newcomer. The company has been around since 2016, previously specialising in applying AI to high-security environments across government, intelligence, defence and law enforcement. In 2020 it raised $3 million from Main Sequence, the venture fund backed by Australia's national science agency CSIRO, and it is now aiming its models at general users as well.

Money remains the constraint, and the company is trying to fix it. Gately said Trellis will go to market imminently for an eight-figure raise, with several investors already interested, possibly before Christmas or early in the new year. Compared with US rivals, he argued, Trellis needs nowhere near as much capital.

On product strategy, Agentaus is proprietary, though Gately said an open-weights version is under consideration. He also said a ranking due to be published on MMLU-Pro will make Agentaus 1 Australia's best performing general AI model, and that he expects naysayers to take shots at any attempt to build a globally relevant model in Australia.

The scepticism is real. Will Liang, founder of advisory firm Amplify AI, has not tested Agentaus 1 but said it will be very difficult for Trellis to make significant commercial inroads: AWS, Google and Microsoft already support Australian-based inference for selected international models, and organisations can run open-weight models, including Chinese ones, on Australian-controlled infrastructure, which weakens data sovereignty as a differentiator. The exceptions, he said, are defence, intelligence and highly sensitive government work, where buyers pay a premium for not depending on foreign providers.

That argument has local history behind it. Earlier this year Washington restricted international access to Anthropic's newest models over their cybersecurity capabilities, prompting calls for Australian-built alternatives; Employment Hero co-founder Ben Thompson compared data dependence to oil supply through the Strait of Hormuz. For Trellis, the test is now concrete: whether 220 billion parameters, local deployment and a proprietary-or-maybe-open licence can win real orders on cost and compliance.