Models that fit your data rules.

You decide where your data is processed. Open-weight models such as Qwen, Mistral and gpt-oss run on the appliance, and your data stays on the device. Hosting in Germany adds larger models, and in On Demand external ones too, depending on your plan.

Free for 3 days, hosted in Germany

Available modelsby deployment type

up to 51models, hosted in Germany

  • Open-weight out of the box · 22included
  • Larger open-weight in hosting · 4included
  • External models (EU/ZDR) · 25depends on plan

As of October 2026, 51 models in the catalogue.

Three locations, one catalogue.

22 models on the appliance

Open-weight models out of the box, for chat, code and embeddings.

Appliance, 22 modelsExample
Choose models for your tasksReady

4 more hosted in Germany

Larger models from Mistral AI, with no hardware of your own.

Hosted in Germany, 4 moreExample
Use Mistral Large 3Ready

25 external models

In On Demand depending on your plan, processed in the providers' EU tenants without storing your input. Not connected with Air-Gap.

25 external modelsExample
Filter by deployment typeReady

When model choice runs into data protection

TodayWith souveraen.ai
Public AI services are ruled out for confidential documents from the start.22 open-weight models run on the appliance, and data stays on the device.
A provider changes its model, and the answers suddenly come out differently.A new version only replaces the old one once you have approved the update.
Large models fail on the hardware that would have to be bought for them.Hosting in Germany adds 4 larger models from Mistral AI.

Model decisions across the company

Data protectionChecks which models fit the company's data rulesWith Air-Gap, external models are not connected, so the question does not arise.
IT managementDecides when a new model version replaces the old oneThe current version keeps running unchanged until approval.
DevelopmentProposes an open-weight model with a commercial licenceAfter testing on the reference appliance, it is added to your profile.
Executive managementDecides whether to add external models on Private SparkThe document base is not transferred, and the provider's token costs apply.

How a model enters your operation

  1. Set where data is processed

    You decide: the appliance, hosting in Germany or, in On Demand depending on plan, external models too.

  2. Select the models

    You decide which model answers, on the appliance according to tasks and concurrent users.

  3. Approve updates

    Models stay in a fixed version, and new ones arrive as tested packages that apply after your approval.

All 51 models in the catalogue.

51 of 51 models

ModelTypeContext windowMax. outputDeployment
Mistral Small 4Mistral AIChat256,000–Local
Ministral 3 8BMistral AIChat256,000–Local
Devstral Small 2Mistral AICode256,000–Local
Magistral Small 1.2Mistral AIChat128,000–Local
Qwen3.6-35B-A3BQwen (Alibaba)Chat256,000–Local
Qwen3.6-27BQwen (Alibaba)Chat256,000–Local
Qwen3-Coder-30B-A3BQwen (Alibaba)Code262,144–Local
Qwen3-30B-A3B-2507Qwen (Alibaba)Chat262,144–Local
Qwen3-14BQwen (Alibaba)Chat32,768–Local
Qwen3-8BQwen (Alibaba)Chat32,768–Local
Qwen3 Embedding 8BQwen (Alibaba)Embedding––Local
Gemma 4GoogleChat128,000–Local
Gemma 3 27BGoogleChat128,000–Local
gpt-oss-120bOpenAIChat131,072131,072Local
gpt-oss-20bOpenAIChat131,072131,072Local
Llama 4 ScoutMetaChat10,000,000–Local
Llama 3.3 70BMetaChat128,000–Local
Llama 3.1 8BMetaChat128,000–Local
Phi-4MicrosoftChat16,384–Local
Phi-4-reasoning-plusMicrosoftChat32,768–Local
Llama 3.3 Nemotron Super 49BNVIDIAChat128,000–Local
Nemotron Nano 9B v2NVIDIAChat128,000–Local
Mistral Large 3Mistral AIChat256,000–Hosted in Germany
Mistral Medium 3.5Mistral AIChat256,000–Hosted in Germany
CodestralMistral AICode256,000–Hosted in Germany
Mistral EmbedMistral AIEmbedding––Hosted in Germany
DeepSeek V4-ProDeepSeekChat1,000,000–External · EU/ZDR
DeepSeek V4-FlashDeepSeekChat1,000,000–External · EU/ZDR
Claude Fable 5.1AnthropicChat1,000,000128,000External · EU/ZDR
Claude Fable 5AnthropicChat1,000,000128,000External · EU/ZDR
Claude Opus 5.5AnthropicChat1,000,000128,000External · EU/ZDR
Claude Opus 5AnthropicChat1,000,000128,000External · EU/ZDR
Claude Opus 4.8AnthropicChat1,000,000128,000External · EU/ZDR
Claude Sonnet 5AnthropicChat1,000,000128,000External · EU/ZDR
Claude Haiku 4.5AnthropicChat200,00064,000External · EU/ZDR
Gemini 3.1 ProGoogleChat1,000,00064,000External · EU/ZDR
Gemini 3.5 FlashGoogleChat1,000,00064,000External · EU/ZDR
Gemini 3.1 Flash-LiteGoogleChat1,000,00064,000External · EU/ZDR
Gemini Embedding 2GoogleEmbedding––External · EU/ZDR
GPT-5.6 SolOpenAIChat1,050,000128,000External · EU/ZDR
GPT-5.6 TerraOpenAIChat1,050,000128,000External · EU/ZDR
GPT-5.6 LunaOpenAIChat1,050,000128,000External · EU/ZDR
GPT-5.5OpenAIChat1,050,000128,000External · EU/ZDR
GPT-5.4OpenAIChat1,050,000128,000External · EU/ZDR
GPT-5.4 miniOpenAIChat400,000128,000External · EU/ZDR
GPT-5.1OpenAIChat400,000128,000External · EU/ZDR
GPT-5OpenAIChat400,000128,000External · EU/ZDR
GPT Image 2OpenAIImage––External · EU/ZDR
GLM-5.2Z.ai (GLM)Chat1,000,000128,000External · EU/ZDR
Kimi K2.6Moonshot AI (Kimi)Chat256,000–External · EU/ZDR
MiniMax-M3MiniMaxChat1,000,000–External · EU/ZDR

In On Demand, external models are paid for with AI credits and processed in the providers' EU and zero data retention tenants. On-premise, they can be enabled optionally, with the provider's token costs. Not with Air-Gap Enterprise. As of October 2026.

Request a model

For IT and data protection

  • Fixed versions

    Models run in fixed versions and are not trained further on your content.

  • Updates only after approval

    New versions arrive as tested packages and only replace the old one after your approval.

  • Your own models are possible

    We test an open-weight model with a commercial licence on the reference appliance and add it to your profile.

Frequently asked questions

Question not listed? Call us: +49 3744 365 2202

Which model answers our questions?

You decide. On the appliance, the selection depends on your tasks and the number of concurrent users.

Can we use GPT, Claude or Gemini models?

In On Demand depending on your plan. On Private Spark as an option, with the provider's token costs. Your document base is not transferred.

Test the models with your own questions.

Free for 3 days, then from €24.90 per user per month. Cancellable monthly.

Free for 3 days, then from €24.90 per user per month.Free for 3 days
Try free for 3 days