The Noxtua Model System

Noxtua provides Legal AI that is tailored to the specific needs of European legal professionals. Legal AI that they can trust on every level. This page serves to provide detailed information on the model system and categories so that clients can make informed decisions on which option serves their needs best.

With the new two-tier model strategy and the broadening of the product offering, Noxtua offers clients the freedom to choose the best matching option for their specific needs. Across Europe, the judiciary, public administration, legal departments, and law firms of all kinds and sizes have different requirements and needs. Noxtua caters to all of them.

Sovereignty and resilience

European digital sovereignty is key and Noxtua's two-tier strategy offers resilience and sovereignty through diversification in a European setup. Noxtua is the only Legal AI that offers the option of exclusively hosting the entire system on sovereign European-controlled infrastructure, while now also rolling out the option to access international, closed-source frontier AI models through the European cloud provider Deutsche Telekom. This offers new resilience through diversification in a European setup.

In terms of security and sovereignty, the crucial point is not the provenance of the model but the infrastructure. A model of European origin hosted on US-controlled infrastructure does not meet the criteria of European data sovereignty; a non-European model hosted on European-controlled infrastructure does, because it is under full European control.

Flexibility and freedom of choice

The new model categories give new flexibility in a time of rapidly evolving model development. They define criteria for the base models that stay the same while the specific models might also change as the state of the art evolves, and after we've tested them thoroughly.

Administrators set which model categories and therefore base models are available in a workspace and which is the default. If you don't see a model dropdown, the multi-model system isn't available in your country, or your organization's administrator has limited which model categories are available. Within that list, each user selects the model category they want, for a whole workspace or a single task, and can change it at any time in the model picker.

Closed- and open-source models

With the new model strategy, Noxtua offers now different product versions based on closed- or open-source models. Closed-source models are proprietary models that keep the model’s code and weights (trained parameters) private (example: Gemini). Open-source models can be downloaded, run locally, fine-tuned, and build upon, completely independent from their original creator. Open-source models in the wider sense include open-weights models that release only the weights (trained parameters). However, they can’t be recreated from scratch because the training code and data oftentimes stay private (example: Mistral, GLM). Open-source models in the strict sense also release the training code and data so that they can be recreated from scratch.

Categories in the Noxtua AI system

Categories of base models in the Noxtua AI system

From now on, Noxtua offers different categories with different base models in the AI system. These base models are combined in the Noxtua AI system with an additional set of specialist models for OCR, search, and voice which are hosted on European infrastructure.

Across every category, client data is never used to train the underlying models and is subject to zero data retention during processing, which means that the models keep nothing once a request is finished. The clients’ conversation history, which enables them to return to their work across sessions and interfaces, is stored by us in encrypted storage on our own European infrastructure (Deutsche Telekom and IONOS), under our own keys, and never on a model provider's systems. Clients can decide for how long this history is kept and when to delete it.

Athena Agent (“Standard”)

Minerva Agent (“Frontier”)

Submodels in the Noxtua model system

The OCR, Voice and Search models all run on Noxtua's own sovereign European infrastructure (Deutsche Telekom/T Cloud Public and IONOS), with zero data retention. Noxtua runs local copies of the open-source/weights base model on our own infrastructure. The models are completely disconnected from the original model creator. Therefore, it is technically impossible for the original creator to access the models or any of their in- or outputs.

Transforming texts: OCR Model

Our OCR model turns every uploaded page into structured, searchable text: scans, faxes, exhibits, handwritten notes, stamps and tables. Every document workflow passes through it. Processing takes place on Noxtua's own European infrastructure (T Cloud Public and IONOS).

  • Model: currently LightOnOCR-2-1b (open weights)

  • Creator: LightOn (France) lightonai/LightOnOCR-2-1B · Hugging Face

  • Operation: own infrastructure T Cloud Public and IONOS, confidential

  • Role in the system: active in every workflow, regardless of which category was chosen

Search, embedding and retrieval across legal content: Noxtua Voyage Embed

Noxtua Voyage is our search model. It turns legal text into the numerical coordinates described above, so the system can find the relevant provision, commentary or precedent by meaning rather than by keyword. It is specifically post-trained by Noxtua for legal content and terminology. Processing occurs on Noxtua's own European infrastructure (T Cloud Public and IONOS). 

  • Model: currently Voyage Law, post-trained by Noxtua for legal content

  • Creator: MongoDB (USA) Voyage AI by MongoDB - Voyage AI by MongoDB - MongoDB Docs

  • Operation: own infrastructure T Cloud Public and IONOS, confidential

  • Role in the system: powers search and retrieval in every workflow, regardless of which category was chosen

Voice Input & Dictation: Voice model

Dictation is how a large part of the profession works. The Voice model transcribes speech to text inside the product, on our own infrastructure, so spoken client matters never leave the environment. Users’ voice and any spoken information remain fully confidential. Noxtua only transcribes speech between the moments users start and stop Voice Input, and only once they have granted the necessary permissions. The speech is processed only in temporary memory, never logged or stored anywhere, and is deleted the moment transcription finishes. Processing takes place on Noxtua's own European infrastructure (T Cloud Public and IONOS). 

  • Model: currently Whisper Large v3 Turbo (open weights)

  • Creator: OpenAI (USA) openai/whisper-large-v3 · whisper large v3 turbo

  • Operation: own infrastructure T Cloud Public and IONOS, confidential

  • Role in the system: available in every workflow, regardless of which category was chosen

Operations & security: our contractual framework

For our self-hosted models, the infrastructure provider cannot access client or case data, and the models run with hash-verified weights (a check confirming the model file is genuine and has not been altered), disconnected from the original creator. The models run in our environment with no outbound network path to the original creator, who therefore has no access to the models or to any inputs or outputs.

For Minerva, the closed-source model runs in a dedicated Google environment (Gemini Enterprise Agent Platform) in a EU data center, reached through Deutsche Telekom over an encrypted, internet-isolated channel; processing is transient. Client traffic doesn’t reach Google over the public internet. Telekom is the contractual counterparty and data processor; Google is with its Gemini Enterprise Agent Platform Deutsche Telekom’s sub-processor operating the model. As a result, Google does not maintain a permanent database that could be the subject of a potential request for disclosure.

Across every category, your data is never used to train the underlying models and is subject to zero data retention during processing, which means that the models keep nothing once a request is finished, including the closed-source model on the Minerva path. 

This must be distinguished from your conversation history: in order for you to be able to return to your work across sessions and interfaces, Noxtua stores your conversations in encrypted storage on our own European infrastructure (Deutsche Telekom and IONOS), under our own keys, and never on a model provider's systems. Clients can decide for how long this history is kept and when to delete it. 

The entire chain in both categories is built to satisfy the professional-secrecy obligations of the jurisdictions we operate in (e.g., Section 203 German Criminal Code; Article 321 of the Swiss Criminal Code or the official Austrian guidelines for practicing the legal profession (Section 40 (3) RL-BA 2015).

Noxtua is certified to ISO 27001, ISO 42001, ISO 9001, ISO 27017, and ISO 27018, and holds a BSI C5 certificate, a SOC 2 report, and a TISAX label.

Berlin

Paris

Stockholm

Zagreb

Munich

Fribourg

English
English
English