Frontier-class capability no longer requires training a frontier model from scratch. It can be orchestrated — and we've built a working proof that lifts open models up a full tier. Here are the three paths to an Australian frontier model, and where we start.
The nations that control their own frontier-class AI control a strategic capability — for science, defence, industry and independence from foreign model providers and their export controls. The conventional assumption is that this means spending billions to pretrain a frontier model. It doesn't have to. A second route — orchestration — is now proven to reach frontier-class results, and it's a route an Australian lab can take today.
Ranked from hardest to most achievable. The capability at the end of each is comparable; the cost, time and risk are not.
Build a GPT/Claude/Gemini-class foundation model domestically. This is the "true sovereignty" path — and it's brutally hard: a fleet of 100,000+ accelerators, vast proprietary data, one of the few dozen world-class pretraining teams on Earth, 12–18 month cycles, and a target that keeps moving. A handful of organisations globally can do it.
Train a coordinator model that learns to decompose a task, route sub-tasks across a pool of models (including itself, recursively), and synthesise the result — all behind a single API. A handful of labs have already used this to match frontier benchmarks. Far cheaper than path 1 (no frontier pretraining), but you still train a model — and the result depends on the models in the pool.
A hand-engineered orchestrator: decompose → route each sub-task to the best available model → verify with multi-agent debate / best-of-N → aggregate. No model training — pure systems engineering with today's tooling. It reaches frontier-ish output by coordinating existing models and verifying hard. The quickest route to frontier-class capability.
Before scaling paths 2–3 to frontier, we prove the technique works one layer down: take fairly non-frontier, open-weight models and orchestrate them up to near-frontier performance. A lower starting point, chosen on purpose — so it's cheap to run and clean to verify.
We take open models and push them up to near-frontier capability — one deliberate layer down from the frontier itself. Why a layer down? Because at this level the result is both verifiable and immediately useful, not just a benchmark claim.
We score on a public benchmark (MMLU-Pro) against the objective answer key — no model judging another. The honest finding: across a panel of free open models the correct answer is present 89.7% of the time — a selection ceiling +7.7 points above the best single model, and above a current frontier model. The winning knowledge is provably in the panel; capturing it reliably by coordination is the open frontier. Full method, every number, and the negative results: the research paper.
The proof is the product: a bootstrapped, ~$0 higher-class model built from free / open weights via coordination. A capability we can then use across everything we run — not a slide, a working asset.
A working demonstration that orchestration lifts models a full tier is the credible foundation for the national pitch: scale the same method (paths 2–3) with real resources, and Australia has sovereign frontier-class AI capability.
We don't claim to have pretrained a frontier model. We claim — and demonstrate — that coordination lifts a tier. That's a real, defensible, repeatable result, and the right first step on a serious path.
Accuracy on MMLU-Pro, scored against the objective answer key. The selection ceiling — the score if you always picked the model that's right — sits +7.7 above the best single model, above the frontier target. Simple voting doesn't capture it yet; synthesis is the lever. See the full paper →
The proof-of-concept de-risks the whole thesis. With the technique demonstrated, the path to an Australian frontier-class model is an engineering and resourcing question — not a moonshot. We're building the proof now; the national-scale version is paths 2 and 3 with the backing to match. If you're working on Australian sovereign AI capability and this is your problem too, that's exactly the conversation to have.