Trust center

Control you can prove.

Where your data goes in each run mode, what protects it, and what is built today against what is planned. Written for security and procurement teams.

Request the security questionnaire

The owner’s Key standing in its reader on dark basalt

Illustrative data. An owner can roll back to the last good version in 60 seconds, by design.

Data handling

Where your data goes, mode by mode.

Compared	Eigna Local	Eigna Hybrid	Eigna Cloud
In one line	Data never leaves.	Data stays in the region you choose.	Data stays in the region you choose.
Runs on	Your machine or an Eigna Box on your site	Small experts on your machine; the main expert on our GPU pool in your region	A dedicated GPU we rent and run for you, in your region
Your data	Stays on your machines	Prompts for the main expert travel over mutual TLS, with no prompt logs by default	On your own volumes, with keys you hold
Training	An encrypted snapshot goes to a GPU we rent in a region you approve, and is deleted after the run	In the region you choose; the exact path is confirmed per install	In the region you choose; the exact path is confirmed per install
Limit	The machine is the limit	30 million pool tokens a month per site	The GPU is the limit

Planned GPU regions are named once terms are signed. Moving between modes is planned.

Eigna Local

In one line: Data never leaves.
Runs on: Your machine or an Eigna Box on your site
Your data: Stays on your machines
Training: An encrypted snapshot goes to a GPU we rent in a region you approve, and is deleted after the run
Limit: The machine is the limit

Eigna Hybrid

In one line: Data stays in the region you choose.
Runs on: Small experts on your machine; the main expert on our GPU pool in your region
Your data: Prompts for the main expert travel over mutual TLS, with no prompt logs by default
Training: In the region you choose; the exact path is confirmed per install
Limit: 30 million pool tokens a month per site

Eigna Cloud

In one line: Data stays in the region you choose.
Runs on: A dedicated GPU we rent and run for you, in your region
Your data: On your own volumes, with keys you hold
Training: In the region you choose; the exact path is confirmed per install
Limit: The GPU is the limit

Planned GPU regions are named once terms are signed. Moving between modes is planned.

Controls

Keys, installer, audit, rollback.

Designed means part of the product design and our release method. Planned means not built yet.

DesignedApproval before learningOnly examples a person approves can train.
DesignedThe release gateNo update without a paired win on a sealed test set. A tie keeps the live version.
DesignedRollbackBack to the last good version in 60 seconds, at any time.
DesignedClient separationClient data never moves between clients.
PlannedKeys you holdCloud volumes under your own keys.
PlannedSigned installerInstalled from signed, versioned templates.
PlannedAudit recordEvery request and every release on record.
PlannedAir-gapped profileToday: Local on a closed network.

Runs today: the scoring harness and sealed tests, training on local machines and H200 GPUs, and the published results at benchmarks.eigna.ai.

Model licences

Every base model passes a licence check.

An owned model is only owned if its base may be used that way. We check the licence of every base and teacher model before use.

Compared	Base models we work with	Licence check
Time series	Kronos-small, the base of our first expert	MIT, verified
Language	Mistral Small 4; Qwen3, Llama and gpt-oss models	Mistral Small 4: Apache 2.0, verified. Others checked before use
Decision and routing	Small open models from 0.42B	Checked before use
Voice, image, video	Open models per task	Checked before use
Not ownable	Hosted-only models, and weights licensed for non-commercial use	Not used unless cleared

Teacher models for distillation are used only where their licence allows it.

Time series

Base models we work with: Kronos-small, the base of our first expert
Licence check: MIT, verified

Language

Base models we work with: Mistral Small 4; Qwen3, Llama and gpt-oss models
Licence check: Mistral Small 4: Apache 2.0, verified. Others checked before use

Decision and routing

Base models we work with: Small open models from 0.42B
Licence check: Checked before use

Voice, image, video

Base models we work with: Open models per task
Licence check: Checked before use

Not ownable

Base models we work with: Hosted-only models, and weights licensed for non-commercial use
Licence check: Not used unless cleared

Teacher models for distillation are used only where their licence allows it.

Responsible use

What it will and will not do.

Never weapons targeting: Defense and security work only through cleared partners, scoped to analysis, logistics and administration.
A person approves what it learns: Nothing trains your model until someone on your side approves it.
Refusal training: Experts trained to refuse hacking and similar misuse, with a red-team refusal rate measured at every update. Planned.
Restricted skills: Penetration-testing skills only for certified firms. Planned.

Regulation

The EU AI Act, by date.

Eigna is designed to support your compliance work, not to replace it. A named owner, a Model Deed, sealed tests and a record of every release make that work easier to show.

2 August 2026Transparency duties apply
2 December 2026New prohibitions apply
2 August 2027National AI sandboxes due
2 December 2027High-risk duties for Annex III uses, among them critical infrastructure and essential public services
2 August 2028High-risk duties for Annex I products

EU AI Act as amended by Regulation (EU) 2026/1744. A company that tunes a model is presumed to become its provider only above a third of the original training compute; we check this per case.

Security questionnaire

Send us your questionnaire.

We answer in writing, per run mode, with what is built and what is planned.

Sub-processors

Listed here once signed.

GPU providers and data centres per region are named on this page once their terms are signed. For this website, see our privacy policy.

Privacy