Sovereign AI Models — Built with SCX.ai
Your own AI model. Built for your country, your industry, your data.
We help governments, banks, and large organisations build sovereign AI models — on open weights you own, in your languages, aligned to your regulators. We've already done it for Australia. We can do it for you.
A 30-minute call with an SCX engineer. No sales sequence. No deck.
Why this, why now
The world is splitting into two kinds of AI.
One is rented from a handful of foreign labs. The other is yours — built for the people you serve, the laws you operate under, and the data you can never send offshore.
Models are becoming national infrastructure.
Just as countries once had to decide whether to build their own telecoms networks, payment rails, and identity systems, they now have to decide whether the language their citizens and regulators use every day is interpreted by a model they control. That decision is being made in boardrooms and ministries right now.
Generic models fail on local context.
A model that doesn't know your regulators, your case law, your dialects, or your industry's working language will sound fluent and be wrong. In finance, healthcare, justice, and government, “sounds fluent and is wrong” is not acceptable.
In a recent SCX evaluation, a leading US base model gave jurisdictionally incorrect answers on Australian regulatory questions in 38% of test cases. Project MAGPiE reduced that to under 4%.
The window to act is open — but it won't stay open.
The open-weights model ecosystem has matured to the point that a serious, domain-specific sovereign model is now a realisable project for a serious organisation, not a moonshot. That window will close as more buyers move and as the easiest open-weights bases get dominated by the largest labs.
Open-weights models are now within 2–4% of leading proprietary models on most reasoning benchmarks, at a fraction of the inference cost.
“The question is no longer 'can we afford to build our own sovereign model?' The question is 'can we afford not to?'”
— David Keane, Co-Founder & CEO, SCX.ai
What we mean by a sovereign model.
A sovereign model is one that an organisation (or a country) controls end-to-end: the weights, the data it was trained on, the evaluation that proves it works, and the right to run it wherever it needs to run.
Built on open weights.
The model starts from openly published base weights (such as GPT-OSS, Qwen, DeepSeek, Llama, or Mistral). You are never renting a black box — you are extending a foundation you could, in principle, rebuild without us.
Trained on your context.
Your data, your language, your regulators, your industry's working materials. The model is shaped by the world it has to operate in, not the average of the open internet.
Evaluated against your success criteria.
Before delivery, the model is tested on a domain-specific evaluation suite — accuracy, bias, jailbreak resistance, regulator-style scenarios — that you define and can re-run.
Yours to operate, wherever you choose.
The model can run on infrastructure you control, on SCX.ai infrastructure, or on a partner's. Sovereignty is a property of the model, not of the data centre it happens to sit in this quarter.
How SCX.ai helps you build a sovereign model.
We work with your team to design, build, evaluate, and hand over a sovereign model that fits the way your organisation actually operates. The result is yours: weights, data, evaluation suite, runbook, and a team that knows how to operate it.
We help you define what to build.
We start with a Model Brief — a working document that defines the model's purpose, the use cases it must serve on day one, the languages and regulators it must respect, and the success metrics it will be measured against. Your experts, our engineers, one signed brief.
We help you prepare the data and train the model.
We work with you to clean, structure, license-check, and split the data the model will learn from. We pick the right open-weights base for the job and fine-tune, align, and (where appropriate) distill a model on it. You see sample completions weekly. Nothing about your data trains anyone else's model.
We help you evaluate it, ship it, and own it.
We build a domain-specific evaluation suite — accuracy, bias, jailbreak resistance, regulator-style scenarios — and run it independently. We deploy the model to wherever you need it and run a structured handover so your team can operate and evolve the model without us.
At the end of the engagement, what we built is yours. There is no “SCX model” hiding under the hood. There is yours.
An engagement, not a product. We start with a conversation, not a contract.
What we've already built.
Built and maintained by SCX.ai
Project MAGPiE — a sovereign reasoning model for all Australians.
A reasoning model we built, in a public-good posture, for use across Australian government, financial services, and the legal sector. Fine- tuned on curated Australian legal, regulatory, and cultural corpora, evaluated against ASIC, APRA, and Privacy Act scenarios, with auditable reasoning traces for compliance teams. In production since early 2026.
- 30–40% more efficient than prompt-engineered alternatives on the same hardware
- Jurisdictionally accurate answers where the underlying base model gave wrong answers in over a third of test cases
- A drop-in OpenAI-compatible API — no rewrites of customer tooling
Built and maintained by SCX.ai
SCX-coder — a high-performance coding model for agentic work.
A coding model we built for software engineering teams and agentic-AI platforms. Designed not just to autocomplete code, but to plan, execute, test, and refine across real engineering environments. Distilled and aligned on an open-weights base, optimised for sustained agentic workloads.
- Up to 6× higher throughput than leading proprietary coding models on agentic tasks
- ~41% lower end-to-end latency on agentic workflows
- 80–90% lower cost per task on high-volume coding
Sovereign AI for New Zealand
Kererū Partnership — Sovereign AI Infrastructure for Aotearoa
SCX.ai is partnering with NZ consortium Kereru.ai to build New Zealand's first fully sovereign AI platform. Powered by SCX's purpose-built inference architecture, the partnership delivers local AI infrastructure with ~70% energy savings over GPU-based systems and includes Project Kererū — a sovereign NZ language model supporting Te Reo Māori, local regulatory contexts, and sector-specific capabilities for government, research, and regulated industries.
MAGPiE, SCX-coder, and Project Kererū each started as a build brief. The next one could be yours.
Who this is for.
Governments and public-service agencies
Ministries, regulators, statistical agencies, courts, central banks, and sovereign-cloud initiatives that need a model aligned to their country's laws, languages, and citizens — and that can show their auditor how the model was built.
What you'd build first: A bilingual or national-language sovereign reasoning model aligned to your most-regulated use case (tax, justice, social services, statistics).
Banks, insurers, and other regulated enterprises
Institutions that handle sensitive customer data, work under strict regulators, and need a model that can be fine-tuned on proprietary corpus (credit policy, claims, KYC, contracts, transactions) without that data leaving their control.
What you'd build first: A domain-specialist model fine-tuned on your internal corpus, with a regulator-ready evaluation suite and a documented build manifest.
Telcos, healthcare networks, utilities, and large engineering organisations
Operators with high-volume, domain-specific workloads — network operations, clinical coding, grid forecasting, agentic software engineering — where a general model is too slow, too expensive, or not accurate enough.
What you'd build first: A specialist model fine-tuned on your operational data, evaluated against your real tasks, and integrated into your existing tooling via an OpenAI-compatible API.
If what you need is a generic chatbot or a one-off prompt-engineering exercise, we can introduce you to a partner who is better suited. The build service exists for the cases that need a real model.
What the journey looks like.
Most clients start with a conversation. A few go straight to a pilot. Almost all do a pilot before a full build.
Step 1
A conversation.
We listen. We tell you honestly whether a sovereign model is the right answer for your situation, and if it isn't, who you should be talking to. If it is, we agree on what the next step is.
Step 2
A pilot (if you want one).
A short, fixed-scope engagement where we write a Model Brief, do a data-readiness assessment, run a feasibility fine-tune on a small slice of your data, and deliver a written recommendation. You walk away with a document, not a debt.
Step 3
A build.
The actual model. You own the weights, the data, the evaluation suite, the runbook, and the trained engineers at the end. SCX remains available as one of your support options, not the only one.
Common questions.
Build the model your country or industry actually needs.
A 30-minute call. An SCX engineer. An honest answer about whether a sovereign model is the right path for you.
No sales sequence. No automated follow-ups. The first call is with someone who has actually built one.