Trellis Data launches 220B Agentaus 1 sovereign AI model | Magica
Trellis Data launches 220B Agentaus 1, testing the market for Australian-controlled AI
Editorial Team
••📖6 min read
Canberra-based Trellis Data is offering its 220-billion-parameter Agentaus 1 model to business and government. Its pitch goes beyond keeping data in Australia, but performance, deployment costs and a planned capital raising remain key tests.
Agentaus 1 has 220 billion parameters, about 50% more than its predecessor, according to Trellis Data.
Australian hosting faces competition from locally deployed international models; control of the underlying technology is a separate selling point.
The company reports an 84.4% MMLU-Pro score and plans an eight-figure capital raising, with further evidence needed on performance and costs.
Trellis Data, a Canberra company founded in 2016 that has supplied AI for government, intelligence, defence and law enforcement, has launched its proprietary Agentaus 1 language model for businesses, agencies and individual users. The new model expands its Australian-hosted AI offering, but its commercial test is whether domestic development and support offer enough value alongside international models that can also run locally.
The company describes Agentaus 1 as a proprietary, 220-billion-parameter model, roughly 50% larger than its predecessor. Parameters are the internal settings adjusted during model training; the increase describes model size rather than independently establishing better performance. Trellis says the update improves accuracy, multistep reasoning and the coherence of longer responses while operating on smaller computing clusters, according to its announcement.
The sources differ on the launch date: the company announcement is dated November 10, while a subsequent report dates the release to November 12. Both describe the same model update. Agentaus, the company’s AI assistant and document workflow platform, was already a commercial service: Trellis announced its original launch on May 7.
The account of the upgrade describes the May offering as an adaptation of open-source frameworks and the new core model as Trellis Data’s own proprietary model. That account does not, by itself, establish the provenance of every component or training dataset.
Microsoft is pitching Copilot Workspaces as a shared environment for people and AI agents. Team collaboration is already in early access, but personal Autopilot integration is still coming soon, with final agent pricing unannounced.
Editorial Team
The platform targets work such as turning policy submissions into ministerial briefs with references back to the original documents. Trellis says documents and processing remain in Australian data centres and it does not train its models on customer material, according to the launch coverage.
The company’s original announcement explicitly attached its promise not to reuse prompts, documents or outputs for model training or commercialisation to its Protected Mode. Buyers need to check the mode and contractual terms applying to their workflow. Access spans free, subscription, enterprise and API tiers.
The product also connects to workplace tools including Xero, OneDrive and Outlook, and incorporates Trellis Data’s voice-recognition software so uploaded audio can feed document generation. Trellis describes these capabilities in the product report; its proposition includes document workflows as well as chat. Local implementation support is another part of the company’s pitch.
Data residency is already a competitive market
Michael Gately, Trellis Data’s co-founder and CEO, previously held a senior technology role in a federal agency. He founded Trellis with Rachel Gately in 2016 after an internal algorithm project stalled over management concerns about risk, according to a September interview. That account describes Agentaus as running on a cloud and software stack owned and controlled by Trellis. Its earlier products included speech recognition and translation on networks disconnected from the internet, as described in the company history accompanying the launch. That record helps explain its focus on customers handling sensitive material.
Yet Australian data residency alone does not make an Australian-developed model necessary. Will Liang, founder of advisory firm Amplify AI, argues that Amazon Web Services, Google and Microsoft already provide Australian-based services for selected international models. Organisations can also run international open-weight models on infrastructure they control in Australia, he said in an assessment of the launch.
Liang had not tested Agentaus 1, so his comments address its market position rather than its measured quality. He nevertheless identified a potential market in defence, intelligence and sensitive government work where independent control of the underlying technology—and operation without foreign providers—may command a premium.
That narrows the question for buyers: whether they need locally processed data, or also a domestic developer responsible for the model itself. Agentaus remains proprietary; Gately said the company was considering an open-weights version. Australian development therefore does not yet mean customers can obtain and independently deploy its weights under an announced open release.
Trellis also faces domestic efforts. The same report identifies Sovereign Australia AI’s Ginan and Maincode’s Matilda, making local model development a contested field rather than an exclusive position.
Agentaus 1’s company-reported MMLU-Pro score. The retained coverage does not disclose evaluation settings. Source: Region.
What the benchmark does—and does not—show
Trellis reports an 84.4% score on MMLU-Pro, a test of language understanding and reasoning using multiple-choice questions. It says that placed Agentaus 1 43rd globally and first in Australia at the time of testing. The November 16 report says the company intended to publish those results. Those rankings should therefore be read as company-reported findings, rather than an independently verified overall ranking. The coverage does not identify the full comparison set or evaluation settings behind that placement.
The benchmark’s research paper describes MMLU-Pro as a reasoning-focused question dataset with ten answer options instead of the original MMLU’s four. Its authors found that chain-of-thought reasoning improved performance relative to direct answering. Evaluation settings matter when comparing scores.
A score on that test does not establish the reliability of a ministerial brief, the security of an enterprise deployment or the cost of completing a document workflow. Gately separately argues that Agentaus is sufficient for ordinary professional work even while trailing frontier models; that is a proposition customers can test against their own tasks, rather than a conclusion established by parameter count.
The next tests are capital and deployment economics
Trellis has received less than $10 million in external funding, according to the report on its funding plans, including a $3 million round backed by CSIRO’s venture capital fund Main Sequence in 2020. Gately said it was preparing an eight-figure capital raising, potentially before Christmas or in the new year. An eight-figure round would mean at least $10 million, exceeding its reported external funding to date. It is planned funding, not money already secured. The historical funding total is not a disclosed development budget for Agentaus 1 and cannot establish how cheaply it was built relative to foreign models.
The company argues that newer development methods reduce the capital needed to advance its models. Its smaller-cluster pitch is reinforced by product coverage describing use on commercial computers and servers for organisations hosting it themselves. Those accounts supply no hardware configuration, throughput measurements or comparable cost per completed task to substantiate a saving. Free and paid access tiers establish availability, but do not settle the total cost of a particular enterprise deployment. The launch reports also do not quantify adoption or revenue from the new model.
For prospective buyers, the decisive evidence will be reproducible benchmark results and trials on their own documents, with explicit hosting, access and cost requirements. For Trellis, the pending capital raising will test whether investors will fund continued domestic model development alongside established global alternatives.
Google is rolling out file references and sharing for Gemini skills as adult personal accounts lose the ability to use Gems. Automatic conversion preserves supported material, but file limits and missing tools can interrupt existing workflows.
Muddy Waters disclosed a short position in Nvidia while questioning a possible connection involving senior adviser Ajay K. Puri. Its own account leaves the key identity question unanswered, separating its new claims from an existing server-diversion prosecution.
Nvidia reported $115.8 billion in quarterly revenue in its first earnings release since co-founder Jensen Huang resigned. Strong demand gives interim CEO Colette Kress a financial cushion, while a campaign for Huang’s return collides with unresolved investigations and the board’s reported opposition.
OpenAI launched GPT-6.2 Astra and Sol for professional work, cutting Astra’s standard input and output token prices by 20%. The models offer cheaper paths through selected tasks, but research evaluations, reporting failures and unfinished independent safeguard testing complicate the case for broader autonomy.
Microsoft’s first live speech-to-speech model offers OpenAI API compatibility and planned enterprise integrations. Its published evaluations show a low-cost option, with several rivals completing more agentic tasks.
Perplexity is adding direct collaboration between frontier models to Computer, letting a lead agent coordinate up to 16 supporting agents. Company-reported benchmarks suggest mixed teams can outperform repeated copies of one model, but leave customer costs and production reliability unresolved.
Marsh reported 105 client cyber incidents in the last two weeks of October, against 48 in the final two weeks of May. The broker links the increase partly to AI, while warning that insurance panel access may not guarantee an available response team.
Meta’s Muse personal AI agent has launched in Australia, with some local shopping and booking connections to follow. Its browser and payment tools broaden its reach, while the company’s technical disclosures qualify its privacy promises.
Microsoft and Salesforce are extending their sales software alliance with Copilotforce, combining CRM workflows with Microsoft 365 context under existing permissions. The Copilot plugin enters public preview, while Teams remains in pilot; earlier Microsoft tools and a Claude beta raise the bar for proving its value.
Microsoft expands MAI Image 2.7 from an editing preview into image generation and API access, with cheaper Flash pricing and a Canvas workspace. Benchmark claims depend on the workload, while the MAI Labs interface carries restrictions on commercial use.
OpenAI is widening applications for its legal AI configuration after improving research scores and expanding its search index. Access still requires approval, and benchmark results show substantial limits alongside the gains.
Microsoft has made Word, Excel and PowerPoint creation generally available inside Copilot, combining editable files with workplace permissions and collaboration. The broader shift toward autonomous work remains staggered across previews, while its claimed cost advantage rests on company-funded testing.
Knit Health’s NEA-led Series A funds deployment of Med-BEAM for care navigation and hospital operations. Its early company benchmark beats three Claude models at predicting clinical actions, while better patient outcomes remain unproven by that test.
Science Corporation’s Series D follows its European retinal-implant launch. Clinical results support improved central vision, but U.S. approval, eligible patients and commercial terms remain consequential questions for the startup, reportedly valued at $3.5 billion.
Google will let Gemini users save conversations while opting out of generative AI training on future chats. The global rollout also extends storage of history-off and temporary chats from 72 hours to around two months; human review and longer retention can still apply.
Google is removing the subscription requirement for Gemini’s task mode in six countries, while adding U.S. payment, calling and messaging features. The expansion broadens access to delegated work, but usage limits, permissions and website compatibility still govern what it can finish.
Eligible Pixel owners can join a Play Store beta to generate home-screen widgets from descriptions, but device, language and regional restrictions still apply.
Microsoft’s new in-house coding models compete with outside suppliers inside GitHub Copilot. Company tests show lower task costs at promotional prices, while its continuing financial ties to OpenAI complicate the rivalry.
Valkryn disclosed a $15 million seed round and early military testing contracts. Its evaluator-focused Battle Lab faces existing commercial tools, public testing capabilities and funding constraints that software alone cannot resolve.