DEFENCE · CLASSIFIED

Military & Defence AI: Inference Inside the Enclave, Including at the Edge

Capability that stops working when the satellite link drops is not capability — it is a demonstration.

Written by Jakub Rusinowski · Last updated 2026-08-01 · Hardware figures computed by our VRAM engine

This page is practitioner guidance on deployment mechanics, not legal advice — involve your counsel or DPO for decisions about your specific obligations.

Defence AI deployment means running models inside the accredited enclave that already holds the data, rather than sending that data to a commercial API. For most defence organisations this resolves to open-weight models on accredited on-premise or government-cloud infrastructure at the appropriate impact level, plus a genuinely disconnected edge tier for platforms and forward elements that cannot assume a link. The technical work is unremarkable; the difficulty is accreditation, provenance, and export-control awareness in model sourcing.

Why this is hard

What the deployment gives you

Inference inside the accredited boundary

The model runs at the impact level where the data already lives, so no new cross-domain flow is created. The AI system inherits the enclave's existing accreditation posture instead of opening a fresh boundary discussion.

Edge inference that assumes no link

A compact model on a ruggedised workstation or embedded GPU keeps working when the connection is degraded, denied, or intentionally silent. Sizing follows the platform power and thermal envelope, not the datacentre catalogue.

Provenance recorded, not assumed

Weights hashed, scanned, and version-pinned on the connected side, with the record following the artefact into the enclave. Publisher jurisdiction, licence terms, and export-control posture reviewed before anything is staged.

Defence-in-depth alignment

Identity from the enclave directory, least privilege on the retrieval corpus, full audit of prompts and completions, and no egress from the inference host — controls your assessors already recognise, applied to one more system.

Two tiers: the enclave and the edge

Defence deployments almost always need both, and they are sized very differently. The enclave tier is a normal GPU deployment inside an accredited boundary. The edge tier is a much smaller model on constrained hardware whose defining requirement is that it never expects a network.

Two tiers: the enclave and the edgeAccredited enclave / disconnected platform. 1. Vetted model source: Publisher and licence reviewed; weights hashed and scanned before staging. 2. Enclave registry: Version-pinned artefacts at the appropriate impact level. Sole install source. 3. Enclave inference: GPU serving for analysts and staff, authenticated to the enclave directory. 4. Edge node: Compact model on ruggedised or embedded hardware. Assumes no link, ever.ACCREDITED ENCLAVE / DISCONNECTED PLATFORMVetted model sourcePublisher and licencereviewed; weights hashedand scanned before staging.Enclave registryVersion-pinned artefacts atthe appropriate impactlevel. Sole install source.Enclave inferenceGPU serving for analystsand staff, authenticated tothe enclave directory.Edge nodeCompact model on ruggedisedor embedded hardware.Assumes no link, ever.

What you have to satisfy

Models that fit this deployment

ModelVRAM (Q4)Runs onContextLicense
Qwen 3 14B
Edge tier — 12 GB, ruggedised — Fits an embedded or mobile GPU envelope while still handling summarisation and structured extraction.
ollama pull qwen3:14b
10 GB12 GB GPU (RTX 3060 12GB / 4070)
Mac: 16 GB unified
128,000Apache 2.0
Qwen 3 32B
Enclave workstation tier — 24 GB — Analyst-facing assistant on a single card; permissive licence simplifies the legal review.
ollama pull qwen3:32b
20 GB24 GB GPU (RTX 3090/4090)
Mac: 32 GB unified
128,000Apache 2.0
GPT-oss 120B
Enclave server tier — 2×48 GB — Where the enclave has to match commercial frontier capability for staff work.
ollama pull gpt-oss:120b
73.3 GB2×48 GB GPUs / big unified memory
Mac: 128 GB unified
125KApache-2.0

Deploying AI inside an accredited or disconnected environment?

We help defence organisations and their suppliers architect AI that runs where the data already is — enclave sizing, edge hardware selection within real power and thermal envelopes, model provenance review, and the documentation your accreditation process expects. Vendor-neutral and practitioner-level.

Ask about a workshop →

Frequently asked questions

Can defence organisations use commercial AI APIs?
For unclassified, non-sensitive administrative work, sometimes — several providers hold government authorisations at the lower impact levels. For mission data, controlled unclassified information, or anything classified, the answer is generally no: the data cannot leave the accredited boundary, and an authorisation held by the provider does not transfer to your use case. That constraint, not model quality, is what pushes defence deployments toward open weights on infrastructure inside the enclave.
What do IL5 and IL6 mean for an AI deployment?
They are impact levels from the DoD Cloud Computing Security Requirements Guide that classify the sensitivity of the information a system handles. IL5 covers controlled unclassified information including national-security systems; IL6 covers classified information up to SECRET. For an AI deployment they determine where the inference may physically run, which providers or facilities are eligible, who may administer the system, and how the boundary is monitored. You deploy at the level of the data the model will touch, including anything reachable through retrieval.
How do you run AI at the tactical edge with no connectivity?
By putting a small model on hardware that goes with the platform and never expecting a link. A 12–14B-class model quantised to 4-bit fits a 12 GB embedded or mobile GPU and handles summarisation, structured extraction, translation, and question-answering over a local corpus. The engineering constraints are power draw, thermal envelope, and physical robustness rather than benchmark scores — and the operational constraint is that the model, its corpus, and its updates all have to be staged before deployment.
Do export controls affect which models we can use?
They can, and this is worth raising early rather than late. The model file itself is generally published openly, but your fine-tuning data, prompts, and outputs may be controlled technical data, and controls also reach who may access the environment — foreign-national access to an enclave holding controlled data is a licensing question. Some open model licences additionally carry acceptable-use terms that interact awkwardly with defence work. Have your export-control officer review model sourcing and enclave access alongside the licence, before the programme commits.
Is an open-weight model secure enough for classified work?
The weights being public is not itself a vulnerability — the security property comes from the environment, not from the model being secret. What does require care is the artefact's integrity and provenance: verify hashes independently, prefer safetensors over pickle-based formats since loading the latter executes code, scan before staging, and pin versions. Treat the model as a third-party dependency entering an enclave, apply the supply-chain controls you already have for that, and the residual risk is comparable to any other imported software component.

Go deeper