Sovereign AI for
decision advantage
Executive summary
Australia's 2024 National Defence Strategy identifies decision advantage — the ability to gather, process, and act on information faster than adversaries — as a core capability requirement. Achieving this requires AI systems that operate on classified networks, understand military domain context, and can be trusted by warfighters as team members.
Mistral AI is the only frontier-class model provider that offers open weights, sovereign deployment, and a full-lifecycle training platform — from API to fine-tuning to custom pre-training — under a single roof. This combination enables Defence to build AI capability that is sovereign to Australia, adapted to ADF operational context, and interoperable with AUKUS partners.
Sovereign by architecture
Open weights deploy on Defence infrastructure. No data leaves the network. No dependency on foreign APIs. No vendor kill switch.
Domain-adapted for Defence
Fine-tune on military doctrine, operational terminology, and institutional knowledge. The model reasons like a Defence professional — not a generic chatbot.
Edge to enterprise
From 3B-parameter models on autonomous platforms to 675B-parameter models for intelligence fusion. One architecture across the full deployment spectrum.
Strategic alignment with Defence priorities
Mistral's capabilities map directly to the priorities articulated in the 2024 National Defence Strategy, AUKUS Pillar II, and the Defence AI Centre's stated objectives.
Decision advantage
The NDS identifies decision advantage as a core capability requirement. Domain-adapted models that understand military context compress the time from intelligence to decision — enabling faster, more informed action under uncertainty.
Autonomous systems
DSTG's Trusted Autonomous Systems program, Ghost Bat, and Ghost Shark all require AI that runs at the edge without connectivity. Mistral's small models deploy on autonomous platforms in denied environments.
Allied interoperability
AUKUS Pillar II requires AI that can be shared between allies without sharing classified training data. Open-weight models solve this architecturally — each nation adapts the same base model to its own context.
Sovereign capability
CIO Crozier has publicly emphasised sovereign digital infrastructure. Open weights deploy on Defence-controlled infrastructure — no foreign API dependency, no vendor kill switch, no data leaving the network.
Force structure transformation
AI agents integrated into military workflows will reshape how headquarters staff, intelligence analysts, and logistics planners operate — enabling smaller, more distributed forces to achieve greater effect.
Responsible AI
Australia's 2026 Defence AI policy requires AI that is lawful, accountable, and auditable. Open-weight models enable full inspection of model behaviour — a transparency that closed-source systems cannot provide.
Why domain adaptation matters
Generic frontier models consistently plateau below acceptable performance thresholds on defence-specific tasks. The gap doesn't close with scale — it closes with domain-specific training data.
HTX Phoenix — public safety LLM
Generic LLMs failed on Singapore's multilingual operational context (Singlish, Mandarin, Malay, Tamil). Mistral co-developed Phoenix — a domain-adapted model pre-trained on Home Team corpora across 10 languages, deployed on government-controlled infrastructure.
TRACLM — military domain LLM
Off-the-shelf LLMs demonstrated sub-optimal performance on Army use cases due to domain-specific vocabulary and jargon. TRAC built three generations of domain-adapted models that markedly improved on every military-specific benchmark.
Medical coding — RAG's sharpest failure
GPT-4 achieved only 46% accuracy on medical code generation. Fine-tuned models reached 97–98% — a 25× improvement that no amount of prompt engineering or RAG could approach.
The customisation ladder
Mistral is the only provider that offers every level of model customisation — from API calls to full custom pre-training, autonomous agent orchestration, and embodied physical AI — under a single roof. Start where you are. Scale when the evidence justifies it.
AUKUS alignment
AUKUS Pillar II's AI and Autonomy workstream creates a specific requirement that Mistral's architecture is uniquely positioned to meet: AI models that can be shared between allies without sharing classified training data.
Open-weight models solve this architecturally. Australia fine-tunes on Australian classified intelligence. The UK fine-tunes the same base model on UK intelligence. Both resulting models are interoperable at the architecture level without either nation exposing its data. A closed API model structurally cannot do this.
Mistral's existing defence partnerships in Singapore (HTX, DSO, DSTA) — a close Australian defence and intelligence partner — provide a proven reference for this model within the Five Eyes-adjacent ecosystem.
Use case mapping
Defence's active programs and stated capability gaps map directly to Mistral's product architecture. Each row connects a specific Defence initiative to the Mistral capability designed to address it.
Trusted Autonomous Systems
The TAS program requires trusted and effective cooperation between humans and machines. Trust is not built through accuracy alone — it requires AI that reasons in contextually appropriate ways, communicating in the language and frameworks warfighters recognise. Domain-adapted models achieve this by internalising military doctrine and operational terminology during training, producing outputs that feel like a team member rather than a foreign tool.
Ghost Bat & Ghost Shark
Autonomous platforms operating in denied environments require AI that runs at the edge without connectivity to cloud infrastructure. Mistral's Ministral 3B and 8B models deploy on Jetson and equivalent edge hardware, air-gapped from external networks. Domain adaptation via Forge ensures these models understand ADF-specific mission context before deployment — not after.
Defence AI Research Network
DSTG is building a national AI research community through the Defence AI Research Network. Mistral's engagement model is purpose-built for this — co-development partnerships where Mistral's research engineers work alongside national AI teams to build domain-adapted models. This is the model already proven with HTX and DSO in Singapore, where Mistral embedded with government researchers to co-develop Phoenix.
Ethics, assurance & responsible AI
Defence's responsible AI policy requires the ability to inspect, audit, and explain AI behaviour — particularly for systems that inform operational decisions. Open-weight models enable full model inspection at every layer: weights, activations, and decision pathways. This is a structural transparency that closed-source API models cannot provide, regardless of the documentation or assurance frameworks they offer.
Defence AI Centre milestones
The Defence AI Centre has identified AI that can suggest schemes of manoeuvre and generate coherent response options as a target capability. This requires models that understand military planning frameworks, force structure, and operational terminology at a deep level — not models that pattern-match on internet text. Domain-adapted models trained on Defence planning corpora can reason within these frameworks natively.
Success with Mistral
Mistral AI is trusted by departments of defence, public safety agencies, and sovereign institutions across multiple countries. These partnerships demonstrate a consistent pattern: open-weight, domain-adapted models deployed on sovereign infrastructure.
Proposed next steps
Questions
When your adversaries are training models on their own military doctrine and operational data, what is the cost of Australia relying on generic models trained predominantly on English-language internet text?
DSTG evaluates technologies against scientific rigour. What does it mean that a fine-tuned model achieved 98% accuracy on a task where the best generic model plateaued at 46% — and no amount of prompt engineering closed the gap?
Your analysts are already drowning in data. If an AI agent misclassifies a single intelligence product on a classified network — with no human in the loop — what's the operational consequence of that model never having seen ADF terminology?
Five Eyes partners are already fine-tuning models on their own classified corpora. If Australia doesn't build the same sovereign capability, does that create a dependency — or a gap — in allied interoperability?