Sovereignty
air-gap · on-premise
RequiresClassified information demands air-gapped, on-premise processing with no external dependency.
HelmcodeOpen weights deployed inside your own perimeter, with no call to any external service.
industries · Defense & govtech
AI in classified environments: air-gapped, open weights, zero dependencies
Maximum sovereignty for sensitive and classified work: open weights deployed fully on-premise or air-gapped, with no data touching third-party infrastructure.
compliance
Compliance, built into the stack.
Every regulatory demand mapped to a platform capability that ships built in, with nothing to configure.
No third parties
no data to third-party infrastructure
RequiresNo data can leave to a provider’s cloud, however trusted.
HelmcodeThe full inference stack runs on hardware you control; there is no dependency to sever.
Auditability
open, inspectable weights
RequiresYou must know exactly what runs and be able to inspect it.
HelmcodeOpen weights and published model cards: the system is inspectable end to end.
This page is an informational overview, not legal advice. For your obligations and the risk classification of each system, consult qualified legal counsel. AI Act Guide →
where the AI Act stops
What the AI Act excludes, and how narrowly.
The AI Act does not govern defence. That sounds like relief until you read how the exclusion is drawn: it hangs on the purpose of the use, not on the system, and it can be lost by a change of use. Which makes segmentation an architectural matter.
01
The word is exclusively
Article 2(3) leaves out AI systems used exclusively for military, defence or national security purposes. Exclusively is the whole clause: a system doing that work and something else besides is not covered by the exclusion.
02
It does not matter who you are
The exclusion applies regardless of the type of entity carrying out those activities. A private contractor is in the same position as a ministry, which cuts both ways: no shelter for a supplier whose system is also doing civilian work.
03
A change of use changes the regime
Recital 24 spells it out: a system for those purposes that is used, temporarily or permanently, for others, civilian, humanitarian, law enforcement or public security, falls within the scope. The regime follows the deployment, so being able to keep deployments genuinely separate is worth designing for.
EU AI Act · Regulation (EU) 2024/1689 Regulation (EU) 2024/1689 of 13 June 2024, Article 2(3) and recital 24. The exclusion rests on Article 4(2) of the Treaty on European Union and on national security remaining a Member State responsibility. What governs a given programme instead is national, and is set by your contracting authority. read the report →
use cases
Your most common use cases.
The cases with the most traction in the sector, each with its own page in detail.
RAG over internal knowledge Grounded answers over internal documentation, fully inside the perimeter. Document analysis Analysis over large document sets, air-gapped. Translation Translation without a single word leaving your network. Research Open-source analysis and technical review inside the perimeter, with every claim traceable to its source. Summarization Long reports condensed for a briefing, air-gapped from anything external.
Recommended open models.
A starting point per task type. The full guide maps 80 cases to the open model for each one.
- GLM-5.2
- Top open reasoning, deployable on-premise given the hardware.
- DeepSeek V4 Flash
- High-volume workhorse for air-gapped deployments.
- Gemma 4 12B
- Runs on modest hardware at the edge, inside the perimeter.
We are distilling and quantizing these open models into small, tightly specialised versions, trained for one task rather than for all of them. A model like that runs on less hardware, answers faster and fits where the big one does not, your own datacenter included. If you have a process with volume and stable criteria, that is the conversation we want to have with you.
// faq
Questions, answered.
What the sector's technical, compliance and business teams ask.
Can it run fully air-gapped?
Yes. Open weights deploy inside your own perimeter with no call to any external service; the full inference stack runs on hardware you control.
Which models can we deploy?
It depends on the available hardware: from Gemma 4 on a single GPU at the edge to GLM-5.2 on a cluster. The architecture, not a fixed catalog, drives the choice.
How do we start?
Through a direct conversation about your environment and constraints. This is not a self-service deployment.
// get started
START BURNING TOKENS
Skip the AI infra work. Deploy your first private inference endpoint today.
Flat rate. EU data. OpenAI API compatible.