Trellis Data says Agentaus 1 is a 220-billion-parameter model built in Australia and hosted entirely on Australian servers, now available to global users through agentaus.com.au. CEO Michael Gately's November 17, 2026 X post went viral on HN as "Australia's own AI model, open to the world."
"Open to the world" means signup access, not open weights. DeepSeek fine-tune rumors spread on social threads as sovereign AI hype meets open-model suspicion — Trellis has not published a model card to settle them.
What Trellis officially announced
The November 10, 2026 newsroom release states:
| Claim | Trellis wording |
|---|---|
| Size | 220B parameters, ~50% larger than prior Trellis model |
| Benchmark | 84.4% MMLU-Pro (Gately X), rank 43 (newsroom) |
| Hosting | Australian servers, encrypted data |
| Access | Free, subscription, enterprise, API tiers |
| Training on user data | Not used for training; deleted chats gone for good (marketing promise) |
Founder Michael Gately framed the launch as proof Australia can be an LLM producer, not only a consumer of US and Chinese models.
What Gately posted on X November 17
Gately's thread (HN-linked) promises:
- 220B parameters and 84.4% MMLU-Pro
- Nothing leaves Australia
- Nothing you type trains the model
- Free entry at agentaus.com.au
Third-party summaries such as AI Frontier Post note no architecture disclosure, no training data provenance, and no independent benchmark reproduction — standard gaps for sovereign LLM marketing.
DeepSeek fine-tune rumor and what Trellis has disclosed
After launch, posts asked whether Agentaus is "just DeepSeek with an Australian flag." The November 10 newsroom post presents locally developed capability but does not name DeepSeek. explainx.ai did not locate a public technical rebuttal with layer maps or tokenizer diff.
Until Trellis ships a model card:
| Question | Status |
|---|---|
| Base architecture | Undisclosed |
| Training compute | Undisclosed |
| Data mix | Undisclosed |
| Independence from Chinese base models | Marketed as local, not independently verified |
Compare DeepSeek V4 Pro posts where weights and APIs are inspectable.
Sovereign AI: what buyers should verify
Sovereign here means jurisdiction and ops, not mathematically provable isolation:
- Contractual data residency clauses.
- Subprocessor list if any foreign API is used for tool calls.
- Key management and government access policies under Australian law.
Government buyers already using Trellis in defence and law enforcement contexts, per the newsroom, will demand higher evidence than startup SaaS checklists.
How Agentaus compares for builders
| Need | Agentaus 1 | Open-weight alternative |
|---|---|---|
| Self-host on your GPU cluster | No public weights | Llama, Qwen, Nemotron checkpoints |
| Compliance narrative for AU data | Core pitch | Requires your AU region deploy |
| Benchmarked coding agent | Unverified in public docs | Qwen SWE claims with harness drama |
If you only need API calls, test latency, tool-use, and JSON reliability on your workload — not leaderboard medians from press releases.
Relation to Trellis 3D products
explainx.ai's Hugging Face TRELLIS LoRA guide covers 3D mesh fine-tuning — a different product line from Agentaus text. Do not confuse TRELLIS.2 assets with Agentaus 1 weights.
Hacker News reception
The HN thread had low points but attracted skepticism that "open to the world" mislabels proprietary hosting. That semantic gap matters for SEO and procurement: call it globally available API, not open source.
Procurement questions Australian enterprises should ask
Before Agentaus replaces OpenAI or Anthropic in a regulated workflow, document:
| Question | Why it matters |
|---|---|
| Subprocessors | Tool-calling may hit US APIs unless disabled. |
| Model version pinning | 220B updates could change compliance artifacts. |
| Fine-tune rumor rebuttal evidence | Denial is not a model card. |
| Deletion guarantees | Marketing "gone for good" needs legal backup in DPA. |
| Incident notification | Sovereign pitches assume local breach response. |
MSIT and defence buyers referenced in Trellis copy will expect ISO and IRAP-style answers not visible on marketing pages alone.
Global access versus data residency
Gately's "open to the world" line targets international signups, while data stays in Australia is the residency pitch. Non-Australian users still send prompts to AU regions — latency and privacy law mapping become your problem.
Compare Anthropic River Bend TPU US-centric builds and K2 Horizon fleet open weights: Agentaus optimizes jurisdiction narrative, not weight inspectability.
Competitive landscape in November 2026
Australia already hosts US cloud models via Azure and AWS regions. Agentaus must beat them on price, latency, or sovereign trust — 84.4% MMLU-Pro alone is insufficient without task-specific evals on government forms, legal summarization, and bilingual EN/zh needs common in Pacific trade.
DeepSeek rumor in depth
Rumors spread because frontier MMLU-Pro scores at 220B resemble distilled or continued-pretrain stories from Chinese base models. Trellis markets local development but did not publish:
- Tokenizer hash comparisons
- Layer-norm statistics versus known checkpoints
- Independent auditor letter
Until one appears, classify "proprietary Australian model" as marketing category, not verified architecture. Hugging Face distillation reporting shows why procurement officers are skeptical.
Developer testing checklist
- Run your RAG corpus through Agentaus API with PII scrubbed.
- Measure p95 latency from your geography, not Canberra-only.
- Compare outputs to Qwen or Llama open weights you can self-host in AU Azure.
- Log refusal behavior on sensitive defence prompts if you are a dual-use shop.
Why HN cared despite low points
HN item 50542843 amplified national pride and semantic honesty debates. US readers misread "open to the world" as MIT license — a recurring 2026 confusion as API launches mimic open-weight language.
Claimed vs verified summary
Claimed by Trellis: 220B size, 84.4% MMLU-Pro (Gately X), rank 43, no user training, AU-only hosting.
Verified by independent primary docs in this run: Newsroom post exists; Gately X post exists; API site exists; no public weights; no architecture disclosure.
Related reading
- DeepSeek V4 Pro launch
- Hugging Face TRELLIS 3D LoRA
- K2 Horizon open model fleet
- RightNow YC kernel open API
- China US model distillation Reuters
- Anthropic direct TPU River Bend
Ten-year sovereign AI context
Canberra has funded applied AI in defence and intelligence for years; Agentaus is the public LLM consumer face of that pipeline. Whether Australia becomes a weight exporter or only a hosting wrapper for foreign math will decide if Gately's producer claim ages well.
Founders outside Australia should still study the launch as template messaging for EU, Canada, and Gulf sovereign cloud offers launching in 2026.
Pricing and tiers (what we know)
Trellis advertises free, subscription, enterprise, and API tiers but does not publish a public price sheet in the newsroom post we used. Treat cost comparisons to GPT-6-class APIs as unknown until agentaus.com.au lists per-million-token tables you can screenshot for procurement.
Enterprise buyers should ask whether volume discounts include on-prem connector fees and support SLAs — standard sovereign upsells.
Failure modes if benchmarks are overstated
If 84.4% MMLU-Pro does not reproduce on Australian English legal text or Indigenous language equity tests, government pilots may stall despite viral X posts. Sovereign marketing raises the reputational cost of bench inflation because taxpayer narratives attach.
Independent evaluators should run Transparent task suites (RAG over bills, multihop forms) rather than single multiple-choice scores.
Integration with existing stacks
Trellis claims Agentaus slots into tools businesses already use. Practically, that means M365 Copilot plugins, ServiceNow, or custom REST — none verified in our primary sources. Pilot one workflow (e.g., contract summarization) before enterprise-wide Copilot replacement.
Update path when Trellis ships Agentaus 2
Gately said "we're already working on what comes next" in the newsroom quote. Subscribers should contract for model version migration rights so fine-tuned internal prompts do not break on silent upgrades.
Media literacy for "sovereign AI" headlines
Reporters will conflate Agentaus with open-source nationalism. Your comms team should keep three terms straight: sovereign hosting, domestic company, open weights. Only the first two are clearly true in Trellis materials today.
How to evaluate Agentaus 1 in a two-week pilot
A short, structured pilot gives procurement teams more signal than any launch post. Use the same prompts across Agentaus 1 and one open-weight baseline you can self-host.
- Week one: run your own document set through both models. Score accuracy, refusals, and JSON validity by hand on at least 100 samples.
- Week one: record p95 latency from your own region at peak and off-peak hours.
- Week two: ask Trellis in writing for its subprocessor list, retention terms, and incident-notification window. Compare the answers with the marketing claims above.
- Week two: repeat the 100-sample test after any model update to check for silent drift.
If Trellis answers the written questions clearly and the pilot scores match your needs, the sovereign-hosting pitch has real value. If the answers stay vague, treat the launch as a hosting offer and not as evidence of a new base model.
Academic and civil society lens
Universities piloting local LLMs for student privacy may choose Agentaus over US APIs — but FERPA-equivalent AU law still requires DPA review. Campus IT should not treat flag icons as compliance.
Why explainx.ai covered both HN candidates together
The Australia open model thread and DeepSeek fine-tune rumor are one product story. Splitting them would duplicate sovereign context.
Audit trail for procurement officers
Save PDF exports of November 2026 newsroom language and X posts before marketing edits retroactively. Sovereign contracts often lock claims at signature date.
Contrarian view on sovereign wrappers
Some HN commenters argue 220B hosted APIs are margin plays on foreign weights with AU branding only. Until weights leak or audit, skepticism is rational — not anti-Australian, pro-evidence.
Benchmark figures remain company-stated until an independent lab publishes reproducible evals on Australian workloads.
Trellis benchmark figures are company-stated until a third party publishes reproducible evals.
