Are our data used to train the models?

No. The models run as pinned versions and are not further trained on your content. What the platform learns from your documents sits as sourced knowledge in your knowledge layer, not in model weights.

Can we add a model that is not in the catalogue?

Yes, if it has open weights and a licence that allows commercial use. We review licence, quality and memory needs on the reference appliance and then add it to your profile.

Can we use GPT, Claude or Gemini models?

In On Demand, yes, depending on the plan. They are processed in the vendors' EU and zero-data-retention (ZDR) tenants; what goes there is the single request with selected excerpts, not your collection. On Private Spark, external models are optional, with the vendor's token costs. Air-Gap Enterprise works without external models.

How do new model versions reach the appliance?

As reviewed packages. A new version goes through the same qualification as the first model and only replaces the old build once you approve it. Nothing is fetched automatically.

Models

Which AI models run for you.

Open models such as Qwen, Mistral and gpt-oss run on your appliance, and your data stays on it. Larger open models are added in hosting in Germany. External models are available in On Demand depending on the plan and are processed in the vendors' EU and zero-data-retention (ZDR) tenants.

Browse the catalogue Compare operating modes

Local, or offline with Air-Gap
Open models out of the box
One interface for every task

Model catalogue

Sort by

Model	Type	Context window	Max output	Deployment
On the appliance, out of the box22
Mistral Small 4Mistral AI	Chat	256,000 tokens	–	Available locally
Ministral 3 8BMistral AI	Chat	256,000 tokens	–	Available locally
Devstral Small 2Mistral AI	Code	256,000 tokens	–	Available locally
Magistral Small 1.2Mistral AI	Chat	128,000 tokens	–	Available locally
Qwen3.6-35B-A3BQwen (Alibaba)	Chat	256,000 tokens	–	Available locally
Qwen3.6-27BQwen (Alibaba)	Chat	256,000 tokens	–	Available locally
Qwen3-Coder-30B-A3BQwen (Alibaba)	Code	262,144 tokens	–	Available locally
Qwen3-30B-A3B-2507Qwen (Alibaba)	Chat	262,144 tokens	–	Available locally
Qwen3-14BQwen (Alibaba)	Chat	32,768 tokens	–	Available locally
Qwen3-8BQwen (Alibaba)	Chat	32,768 tokens	–	Available locally
Qwen3 Embedding 8BQwen (Alibaba)	Embedding	–	–	Available locally
Gemma 4Google	Chat	128,000 tokens	–	Available locally
Gemma 3 27BGoogle	Chat	128,000 tokens	–	Available locally
gpt-oss-120bOpenAI	Chat	131,072 tokens	131,072 tokens	Available locally
gpt-oss-20bOpenAI	Chat	131,072 tokens	131,072 tokens	Available locally
Llama 4 ScoutMeta	Chat	10,000,000 tokens	–	Available locally
Llama 3.3 70BMeta	Chat	128,000 tokens	–	Available locally
Llama 3.1 8BMeta	Chat	128,000 tokens	–	Available locally
Phi-4Microsoft	Chat	16,384 tokens	–	Available locally
Phi-4-reasoning-plusMicrosoft	Chat	32,768 tokens	–	Available locally
Llama 3.3 Nemotron Super 49BNVIDIA	Chat	128,000 tokens	–	Available locally
Nemotron Nano 9B v2NVIDIA	Chat	128,000 tokens	–	Available locally
In addition, hosted in Germany4
Mistral Large 3Mistral AI	Chat	256,000 tokens	–	EU hosted
Mistral Medium 3.5Mistral AI	Chat	256,000 tokens	–	EU hosted
CodestralMistral AI	Code	256,000 tokens	–	EU hosted
Mistral EmbedMistral AI	Embedding	–	–	EU hosted
External models25Available in On Demand, through AI credits. Optional in on-premise installations, with the vendor's token costs. Not on Air-Gap Enterprise.
DeepSeek V4-ProDeepSeek	Chat	1,000,000 tokens	–	External · EU/ZDR
DeepSeek V4-FlashDeepSeek	Chat	1,000,000 tokens	–	External · EU/ZDR
Claude Fable 5.1Anthropic	Chat	1,000,000 tokens	128,000 tokens	External · EU/ZDR
Claude Fable 5Anthropic	Chat	1,000,000 tokens	128,000 tokens	External · EU/ZDR
Claude Opus 5.5Anthropic	Chat	1,000,000 tokens	128,000 tokens	External · EU/ZDR
Claude Opus 5Anthropic	Chat	1,000,000 tokens	128,000 tokens	External · EU/ZDR
Claude Opus 4.8Anthropic	Chat	1,000,000 tokens	128,000 tokens	External · EU/ZDR
Claude Sonnet 5Anthropic	Chat	1,000,000 tokens	128,000 tokens	External · EU/ZDR
Claude Haiku 4.5Anthropic	Chat	200,000 tokens	64,000 tokens	External · EU/ZDR
Gemini 3.1 ProGoogle	Chat	1,000,000 tokens	64,000 tokens	External · EU/ZDR
Gemini 3.5 FlashGoogle	Chat	1,000,000 tokens	64,000 tokens	External · EU/ZDR
Gemini 3.1 Flash-LiteGoogle	Chat	1,000,000 tokens	64,000 tokens	External · EU/ZDR
Gemini Embedding 2Google	Embedding	–	–	External · EU/ZDR
GPT-5.6 SolOpenAI	Chat	1,050,000 tokens	128,000 tokens	External · EU/ZDR
GPT-5.6 TerraOpenAI	Chat	1,050,000 tokens	128,000 tokens	External · EU/ZDR
GPT-5.6 LunaOpenAI	Chat	1,050,000 tokens	128,000 tokens	External · EU/ZDR
GPT-5.5OpenAI	Chat	1,050,000 tokens	128,000 tokens	External · EU/ZDR
GPT-5.4OpenAI	Chat	1,050,000 tokens	128,000 tokens	External · EU/ZDR
GPT-5.4 miniOpenAI	Chat	400,000 tokens	128,000 tokens	External · EU/ZDR
GPT-5.1OpenAI	Chat	400,000 tokens	128,000 tokens	External · EU/ZDR
GPT-5OpenAI	Chat	400,000 tokens	128,000 tokens	External · EU/ZDR
GPT Image 2OpenAI	Image	–	–	External · EU/ZDR
GLM-5.2Z.ai (GLM)	Chat	1,000,000 tokens	128,000 tokens	External · EU/ZDR
Kimi K2.6Moonshot AI (Kimi)	Chat	256,000 tokens	–	External · EU/ZDR
MiniMax-M3MiniMax	Chat	1,000,000 tokens	–	External · EU/ZDR

On the appliance, out of the box

Mistral Small 4
Mistral AI
Context window
256,000 tokens
Max output
–
ChatAvailable locally
Ministral 3 8B
Mistral AI
Context window
256,000 tokens
Max output
–
ChatAvailable locally
Devstral Small 2
Mistral AI
Context window
256,000 tokens
Max output
–
CodeAvailable locally
Magistral Small 1.2
Mistral AI
Context window
128,000 tokens
Max output
–
ChatAvailable locally
Qwen3.6-35B-A3B
Qwen (Alibaba)
Context window
256,000 tokens
Max output
–
ChatAvailable locally
Qwen3.6-27B
Qwen (Alibaba)
Context window
256,000 tokens
Max output
–
ChatAvailable locally
Qwen3-Coder-30B-A3B
Qwen (Alibaba)
Context window
262,144 tokens
Max output
–
CodeAvailable locally
Qwen3-30B-A3B-2507
Qwen (Alibaba)
Context window
262,144 tokens
Max output
–
ChatAvailable locally
Qwen3-14B
Qwen (Alibaba)
Context window
32,768 tokens
Max output
–
ChatAvailable locally
Qwen3-8B
Qwen (Alibaba)
Context window
32,768 tokens
Max output
–
ChatAvailable locally
Qwen3 Embedding 8B
Qwen (Alibaba)
Context window
–
Max output
–
EmbeddingAvailable locally
Gemma 4
Google
Context window
128,000 tokens
Max output
–
ChatAvailable locally
Gemma 3 27B
Google
Context window
128,000 tokens
Max output
–
ChatAvailable locally
gpt-oss-120b
OpenAI
Context window
131,072 tokens
Max output
131,072 tokens
ChatAvailable locally
gpt-oss-20b
OpenAI
Context window
131,072 tokens
Max output
131,072 tokens
ChatAvailable locally
Llama 4 Scout
Meta
Context window
10,000,000 tokens
Max output
–
ChatAvailable locally
Llama 3.3 70B
Meta
Context window
128,000 tokens
Max output
–
ChatAvailable locally
Llama 3.1 8B
Meta
Context window
128,000 tokens
Max output
–
ChatAvailable locally
Phi-4
Microsoft
Context window
16,384 tokens
Max output
–
ChatAvailable locally
Phi-4-reasoning-plus
Microsoft
Context window
32,768 tokens
Max output
–
ChatAvailable locally
Llama 3.3 Nemotron Super 49B
NVIDIA
Context window
128,000 tokens
Max output
–
ChatAvailable locally
Nemotron Nano 9B v2
NVIDIA
Context window
128,000 tokens
Max output
–
ChatAvailable locally

In addition, hosted in Germany

Mistral Large 3
Mistral AI
Context window
256,000 tokens
Max output
–
ChatEU hosted
Mistral Medium 3.5
Mistral AI
Context window
256,000 tokens
Max output
–
ChatEU hosted
Codestral
Mistral AI
Context window
256,000 tokens
Max output
–
CodeEU hosted
Mistral Embed
Mistral AI
Context window
–
Max output
–
EmbeddingEU hosted

External models

Available in On Demand, through AI credits. Optional in on-premise installations, with the vendor's token costs. Not on Air-Gap Enterprise.

DeepSeek V4-Pro
DeepSeek
Context window
1,000,000 tokens
Max output
–
ChatExternal · EU/ZDR
DeepSeek V4-Flash
DeepSeek
Context window
1,000,000 tokens
Max output
–
ChatExternal · EU/ZDR
Claude Fable 5.1
Anthropic
Context window
1,000,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
Claude Fable 5
Anthropic
Context window
1,000,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
Claude Opus 5.5
Anthropic
Context window
1,000,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
Claude Opus 5
Anthropic
Context window
1,000,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
Claude Opus 4.8
Anthropic
Context window
1,000,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
Claude Sonnet 5
Anthropic
Context window
1,000,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
Claude Haiku 4.5
Anthropic
Context window
200,000 tokens
Max output
64,000 tokens
ChatExternal · EU/ZDR
Gemini 3.1 Pro
Google
Context window
1,000,000 tokens
Max output
64,000 tokens
ChatExternal · EU/ZDR
Gemini 3.5 Flash
Google
Context window
1,000,000 tokens
Max output
64,000 tokens
ChatExternal · EU/ZDR
Gemini 3.1 Flash-Lite
Google
Context window
1,000,000 tokens
Max output
64,000 tokens
ChatExternal · EU/ZDR
Gemini Embedding 2
Google
Context window
–
Max output
–
EmbeddingExternal · EU/ZDR
GPT-5.6 Sol
OpenAI
Context window
1,050,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT-5.6 Terra
OpenAI
Context window
1,050,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT-5.6 Luna
OpenAI
Context window
1,050,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT-5.5
OpenAI
Context window
1,050,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT-5.4
OpenAI
Context window
1,050,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT-5.4 mini
OpenAI
Context window
400,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT-5.1
OpenAI
Context window
400,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT-5
OpenAI
Context window
400,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
GPT Image 2
OpenAI
Context window
–
Max output
–
ImageExternal · EU/ZDR
GLM-5.2
Z.ai (GLM)
Context window
1,000,000 tokens
Max output
128,000 tokens
ChatExternal · EU/ZDR
Kimi K2.6
Moonshot AI (Kimi)
Context window
256,000 tokens
Max output
–
ChatExternal · EU/ZDR
MiniMax-M3
MiniMax
Context window
1,000,000 tokens
Max output
–
ChatExternal · EU/ZDR

51 of 51 models

Open-weight models run on the appliance and in On Demand out of the box. Larger open-weight models are available in addition in hosting in Germany. External models from OpenAI, Anthropic, Google, DeepSeek and other vendors are available in On Demand through AI credits and optional on-premise, there with the vendor's token costs. External models are processed in the vendors' EU and zero-data-retention (ZDR) tenants.

Missing a model?

We add further open models to your profile after a licence and quality review. Tell us which model you need.

Request a model

Model operation

How models are run on souveraen.ai.

Fixed model versions, model and processing location shown with every answer, updates only after your approval.

Start for free

The right model per task

Chat, code, embedding and image: open models and, in On Demand, external models in EU and ZDR tenants.

You can see which model answered

Every answer names the model that produced it and where it ran.

Nothing changes overnight

A model is updated only after you approve the update.

Operating routes

First the data path, then the model.

A model name says nothing about where your request is processed. So you choose the operating route first; the models that fit follow from it.

Local on the appliance

Open models such as Qwen3.6, Mistral Small 4, Gemma 4, Llama and gpt-oss run on the DGX appliance in your building. External models are optional, with the vendor's token costs.

Data stays local
No token costs for local models
A fixed monthly price

See Private Spark

On Demand

The same software as on the appliance, in a data centre in Germany. We sign a GDPR data-processing agreement with you. German and European law governs the contract and the operation. External models are processed here in the vendors' EU and zero-data-retention (ZDR) tenants.

GDPR data processing agreement
A separate tenant
Start for free

Plans and pricing

External models

In On Demand, models from OpenAI, Anthropic, Google and other vendors are available depending on the plan. They are processed in the vendors' EU and zero-data-retention (ZDR) tenants. On Private Spark they are optional.

In the vendors' EU and ZDR tenants
On-premise optional, with token costs
Not on Air-Gap Enterprise

Discuss your use

Common questions

What people ask most before they choose a model.

Was Geschäftsführung, IT und Datenschutz vor dem Start wissen wollen.

Lieber persönlich sprechen?

Wir klären Einsatzfall, Datenschutz und Betriebsart in einem kurzen Gespräch, ohne Verkaufsdruck. Telefon (03744) 365 2202.

Beratung anfragen

You decide. On the appliance an open model such as Qwen3.6 or gpt-oss-120b does the work; which one depends on your tasks and how many people work at once, and is recorded in the offer. In On Demand, depending on the plan, external models in the vendors' EU and ZDR tenants are added.

The right profile

Which models belong on your appliance?

Tell us your tasks and how many people work at once. We propose the model profile and record it in the offer.

Start for free Talk about a profile

Name tasks and user count1
Receive a model profile2
Fixed in the quote3

Which AI models run for you.

Model catalogue

How models are run on souveraen.ai.

The right model per task

You can see which model answered

Nothing changes overnight

First the data path, then the model.

Local on the appliance

On Demand

External models

What people ask most before they choose a model.

Lieber persönlich sprechen?

Which model answers our questions?

Are our data used to train the models?

Can we add a model that is not in the catalogue?

Can we use GPT, Claude or Gemini models?

How do new model versions reach the appliance?

Which models belong on your appliance?