Models and hosting
The AI you need. Sovereign or supercharged.
NotHuman doesn’t sell one model: we pick the right one for each task. Open models hosted in Switzerland when your data requires it, frontier models when performance counts. On your server, in a Swiss cloud or elsewhere: you decide.
- Open models hosted in Switzerland or on-prem
- Frontier models when performance counts
- Human approval before anything is sent
Example: each file goes to the engine set by the rule. Payslips and invoices go to Apertus in a Swiss cloud, posts go to a frontier model through an API.
Three words to understand it all
- The model The “brain” that reads and writes: GPT, Claude, Gemini, Mistral, Apertus… Some are closed, others open. See the 11 model families
- Hosting The computer this brain runs on: at your premises, in Switzerland or at the provider. It decides where your data goes. See the five options
- Routing The rule that sends each task to the right model: sensitive work in Switzerland, the rest where quality comes first. Find my setup
The models
The right model for each task. No provider imposed.
NotHuman isn’t tied to any AI provider. For each task, we use the model that fits the task and how sensitive your data is.
Performance
Frontier models from the major AI labs, through an API or your cloud account, under contract.
Ideal for Demanding writing, complex analysis, non-sensitive data.
Sovereign
Open models, hosted where you decide: in Switzerland or on your server.
Ideal for Client data, health, HR, professional secrecy.
Good to know: for your data, what counts is where a model is hosted, not where it comes from. An open model from the US or China can run in Switzerland without sending anything to its maker.
Performance · frontier models
4 familiesFrontier models from the major AI labs, used through an API or your cloud account, under contract. For tasks where quality comes first, when the data allows it.
- United States Performance
GPT
OpenAI
- GPT
- mini
- nano
The models behind ChatGPT. All-rounders for writing, analysis and code.
- United States Performance
Claude
Anthropic
- Opus
- Sonnet
- Haiku
Writing, long-document analysis and coding.
- United States Performance
Gemini
Google
- Pro
- Flash
- Flash-Lite
Built to be multimodal: text, images, audio and video.
- France Performance
Mistral
Mistral AI
- Mistral Large
- Mistral Medium
Models from Mistral AI, a French company, used through its API. Large and Medium also come as open weights.
Open and sovereign
7 familiesModels whose weights, meaning the trained model itself, are published and can be downloaded: we host them in Switzerland or on your server, and your data stays where you decided. “Open weights”: the model can be used freely, under its licence. “Fully open”: its code and training data are published too.
- Switzerland Sovereign
Apertus
EPFL · ETH Zurich · CSCS
- Apertus 70B
- Apertus 8B
The open Swiss language model. Apertus 1.5 (July 2026) also reads images and, experimentally, audio.
Licence Apache 2.0 · download after accepting a usage policy
- United States Sovereign
Llama
Meta
- Llama 3
- Llama 4
Meta’s open-weight model family, from small to very large.
Licence Llama Community License
- France Sovereign
Mistral
Mistral AI
- Mistral Large 4
- Mistral Medium 3.5
- Mistral Small 4
- Ministral 3
Mistral AI’s open models, from very compact (Ministral) to very large (Large).
Licence Apache 2.0 (Small 4, Ministral 3) · modified MIT (Large 4, Medium 3.5)
- China Sovereign
Qwen
Alibaba
- Qwen3.8-27B
- Qwen3.6
Open models in many sizes, with variants for vision and code.
Licence Apache 2.0 for the versions listed
- China Sovereign
DeepSeek
DeepSeek
- DeepSeek-V4 Pro
- DeepSeek-V4 Flash
Large open models with several reasoning levels. With us, self-hosted only: nothing goes through DeepSeek’s own service.
Licence MIT
- United States Sovereign
Gemma
Google
- Gemma 4 31B
- Gemma 4 E4B
Google’s compact open models, built from the same research as Gemini.
Licence Apache 2.0
- United States Sovereign
Phi
Microsoft
- Phi-4
- Phi-4-mini
Microsoft’s small language models: fast and frugal, for focused tasks.
Licence MIT
-
Another model in mind?
We evaluate the open models you prefer, on your cases and against their licence.
Let’s talk
Voice and transcription
1 familyTurning audio into text: meetings, voice notes, dictation. Can run locally.
- United States Sovereign
Whisper
OpenAI
- Whisper large-v3
- Whisper turbo
Multilingual speech-to-text, code and weights published under the MIT licence. Run locally, your recordings stay with you.
Licence MIT
-
Example voice pipeline
- Voice note or meeting
- Whisper transcribes, locally
- A model summarises and lists the tasks
- You approve
With Whisper running locally, the audio stays on your server. Recording a conversation requires the participants’ consent.
Names and trademarks belong to their owners. NotHuman is independent: not a partner of, affiliated with or endorsed by these companies and institutions.
Versions and licences as of 30 October 2026, for guidance only: we check them for each project.
Hosting
Where does the AI run? Five options, your call.
The model matters, but where it runs matters just as much: that’s what decides where your data goes. Each option has its strengths; you choose, we set it up.
From the most control to the most raw power: On your server, Swiss cloud, Your cloud account, Provider APIs. Hybrid combines both, data by data.
-
On your server
The model runs at your premises, on your hardware.
- Ideal for
- Professional secrecy, health or HR data; teams that already have a server room.
- Your data
- Stay on your premises: the AI runs on site.
- Models
- Open models: Apertus, Llama, Mistral, Qwen, Gemma, Phi, Whisper.
- Setup
- A server with a graphics card (GPU), sized to the model. We install it, secure it and look after it.
-
Swiss cloud
The model runs in a data centre in Switzerland.
- Ideal for
- Client and financial data, when you want to stay in Switzerland without buying hardware.
- Your data
- Processed in Switzerland by a Swiss host, under a data processing agreement.
- Models
- Open models: Apertus, Llama, Mistral, Qwen…
- Setup
- No hardware at your premises. Shared or dedicated server, depending on volume.
-
Your cloud account
Your own Azure, AWS or Google Cloud account, in a European region.
- Ideal for
- Companies already with Microsoft, AWS or Google Cloud that want frontier models under their own contract.
- Your data
- In your account. Recent frontier models (GPT, Claude, Gemini) are usually processed there in a European zone: in Europe, not necessarily in Switzerland. A Swiss region exists only for some models, often not the newest: we check model by model.
- Models
- GPT via Azure, Claude via AWS or Google Cloud, Gemini via Google Cloud, or open models. For pros: Azure EU Data Zone, AWS Bedrock “EU” profile, Vertex AI “eu” multi-region.
- Setup
- We deploy into your account; the cloud bill stays in your name.
-
Provider APIs
Direct access to frontier models from OpenAI, Anthropic, Google or Mistral AI.
- Ideal for
- Public or low-sensitivity data, demanding writing, a quick start.
- Your data
- Processed by the provider, often outside Switzerland, under a data processing agreement (DPA); processing in Europe available from some, on conditions.
- Models
- GPT, Claude, Gemini, Mistral Large and Medium, in their latest versions.
- Setup
- The quickest setup: an API key and a contract. Pay as you go.
-
Hybrid
Each piece of data goes to the right engine: sensitive work in Switzerland, the rest to frontier models.
- Ideal for
- SMEs with both sensitive data and tasks where writing quality matters.
- Your data
- Sensitive data stays in Switzerland or on your premises; the rest goes through the API, under contract.
- Models
- Apertus or another open model, plus GPT, Claude or Gemini.
- Setup
- A readable routing rule, written with you, decides which engine handles what.
Configurator
Your AI setup in four questions.
Answer four questions: the suggested setup updates live, with the routing diagram. For pros, the full rule is just underneath.
Suggested setup
Hybrid
Each piece of data goes to the right engine: sensitive work in Switzerland, the rest to frontier models.
Why this choice
Pros
Good to know
Example models
Solid: suggested. Dashed: an alternative to test. The final choice comes after tests on your cases.
In every case, someone on your team approves before anything is sent.
The form opens with a summary of your answers, and no personal data.
For pros: the routing rule
The same decision, written as a readable policy: an example generated from your answers, to be validated together.
Fixed rules, computed in your browser: nothing is sent unless you click the button.
With NotHuman
From choosing the model to going live.
No betting on a model: we test on your documents, then go live where you decided.
-
Free audit
30 minutes to list your tasks and classify your data: public, client, sensitive.
-
Test on your cases
We compare two or three models on a sample of your documents: quality, speed and cost.
-
Go-live
On the hosting you chose, with your approval rules. We aim for a first automation in service 2 to 4 weeks after the audit.
-
Care
Monitoring, a monthly report, and a model switch when a better-suited one comes out, without rebuilding everything.
Frequently asked questions
Which AI model is best?
There isn’t a single one: the right model depends on the task, how sensitive the data is and the volume. A frontier model (GPT, Claude, Gemini…) is often stronger on delicate writing; a smaller open model is often enough to sort or extract, and can run in Switzerland. We compare several models on a sample of your real documents before choosing.
Will our data be used to train the models?
It depends on the provider and the plan. According to their terms, the business plans and APIs of the major providers generally don’t use your data for training by default, unlike many personal accounts. We check each service’s terms, sign the necessary agreements and configure the settings; the commitments remain the provider’s.
With an open model hosted in Switzerland or on your premises, the model doesn’t learn from your requests and your data isn’t passed to its maker. More in our guide ChatGPT at work and data protection.
What is Apertus?
Apertus is the open Swiss language model, built by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS). Its weights, code and training data documentation are published under the Apache 2.0 licence (download after accepting a usage policy). It comes in 8 and 70 billion parameter versions; Apertus 1.5, released in July 2026, also reads images and, experimentally, audio.
It can run in a Swiss data centre or on your servers. Its results differ from those of frontier models, so we test it on your real cases before recommending it.
Can everything run on our own server?
Yes, with open models: Apertus, Llama, Mistral, Qwen, Gemma, Phi, and Whisper for voice. You need a server with a graphics card (GPU) suited to the model’s size: an 8 billion parameter model needs far less computing power than a 70 billion one.
Closed models such as GPT, Claude or Gemini can’t be downloaded: they’re used through an API or your cloud. We size the hardware with you.
Can we switch models later?
Yes. We build model-agnostic pipelines: the model is a setting, not a foundation. When a better-suited model comes out or your needs change, we test it on your cases, then switch without rebuilding the automation.
NotHuman
Not sure which engine? Let’s choose it together.
In a free 30-minute audit, we look at your tasks, your data and the hosting that suits you. You leave with a clear recommendation.