Agentaus 1 is a commercial Australian large language model, sold by Trellis Data on free and paid plans. Founder Michael Gately posted on X on 17 November that the 220-billion-parameter model scores 84.4% on the MMLU-Pro benchmark, ranks 43rd in the world and keeps everything in Australia. The post has passed 1.3 million views. The model isn’t new: Trellis announced it on 10 November and launched it on Thursday 12 November, InnovationAus and Region report.
Trellis’s own release gives the 220 billion parameter count, about 50% more than its previous model, and says Agentaus 1 is available on free, subscription, enterprise and API tiers. Three of the post’s claims need more than the post itself to stand up: the ranking, the data claim and the launch date.

What the ranking rests on
Trellis’s release cites a world ranking of 43 for foundational models and doesn’t name the benchmark. Australian Cyber Security Magazine reports the company tied the ranking to MMLU-Pro. Region reports an 84.4% score that put the model 43rd at the time of testing, in results Trellis will look to publish soon. Trellis hasn’t named the list it ranked against or the test settings, and the rank depends on which models that list counts.
MMLU-Pro is a multiple-choice knowledge test with 10 answer options per question. Kaggle’s MMLU-Pro leaderboard, updated 21 October, has a top score of 93.2%. The AIB tracker, which republishes Artificial Analysis results, lists Anthropic’s Claude Opus 4.5 at 89.5%. An 84.4% score sits nine points below the top of Kaggle’s board.

Where the data claim holds
The post says nothing ever leaves Australia. Agentaus’s FAQ says Protected Mode processes prompts, uploaded documents and outputs on Australian-owned infrastructure in Australian data centres, with nothing routed overseas. The privacy policy says the platform is hosted solely in Australia. Trellis says it runs its own cloud rather than AWS, Azure or other public hyperscalers, and it doesn’t disclose the facility operators or its sub-processors.
The pricing page lists a second option on the Free, Standard and Professional plans: External Mode, described as global engines with results kept protected, paid for with monthly credit. Clause 9.3 of the User Agreement says that when you spend External AI Credits, prompts and model outputs may be transmitted to the third-party provider, and that you remain responsible for any privacy obligations the provider imposes. The post doesn’t say which mode it describes. Trellis’s May launch release tied its promise not to reuse customer data for training to Protected Mode.

The no-training promise itself is written into the User Agreement, which says Trellis never uses customer prompts, responses or documents to train, improve or fine-tune its models. On deletion, the FAQ says deleted chats and files are erased from live systems and backup copies are removed as soon as practical. The privacy policy says Trellis doesn’t retain backup copies after a deletion request.
Which model sits behind Protected Mode
Region reports that the May version of Agentaus adapted open-source frameworks, while Agentaus 1 is Trellis’s own proprietary 220-billion-parameter model. The FAQ says all Agentaus models run in Protected Mode using customised open-source models hosted by Trellis, and that Trellis is building its own frontier model with an announcement expected soon. The FAQ doesn’t mention Agentaus 1 by name, and Trellis hasn’t said whether the page predates the launch.
What launch week showed
Agentaus’s account posted on 17 November that the free tier covers questions about PDFs, Word documents and spreadsheets with answers linked to the source, image analysis and an iPhone app on the Australian and Canadian App Stores. Gately later posted that demand is high and the service is running slower while Trellis adds capacity.
Paid personal plans on the pricing page are A$24 a month for Standard and A$107 a month for Professional. The FAQ says Trellis has held ISO/IEC 27001 certification since 2022, and that its Essential Eight Maturity Level 2 rating is a self-assessment, with a formal independent assessment scheduled within 12 months.
What Trellis hasn’t published
Trellis’s release doesn’t give the model’s architecture or training data. Region reports the company will look to publish the benchmark results soon. Trellis hasn’t given a date.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.




