A more accessible option for local testing, workstations, targeted private assistants and lighter production workloads.
Official model ↗Build private, enterprise-ready AI on Apertus — the fully open foundation model developed in Switzerland. Keep control over the model, infrastructure, data and the way your people use AI.
Switzerland's fully open foundation-model family, developed by researchers at ETH Zurich, EPFL and CSCS within the Swiss AI Initiative. Trained on the Alps supercomputer in Lugano and designed to be inspected, adapted and deployed.
Apertus makes its weights, architecture, training methods, data documentation and code available for inspection and independent development. Its original generation was trained on data spanning more than 1,000 languages; real performance still varies by language and task.
Apertus 1.5 adds image understanding, experimental audio understanding, stronger tool use, instruction following and an optional reasoning mode. It accepts multimodal input and returns text. Training and deployment details are published for technical teams to assess.
*The maximum context length is a model capability, not a guarantee that a specific hardware setup or serving configuration will support it efficiently.
A more accessible option for local testing, workstations, targeted private assistants and lighter production workloads.
Official model ↗The full-scale model for demanding use cases, generally requiring server-class acceleration and careful capacity planning.
Official model ↗Smaller distilled variants from Apertus 1.1 for resource-constrained devices and specialised workflows.
Model collection ↗Apertus is developed by the Swiss AI Initiative, not Pulsanto. Apache 2.0 allows commercial use, but deployment teams should also review the model's usage policy, access conditions, privacy documentation and release-specific limitations.
“Sovereign AI” is not a single product feature. It is the ability to decide where your information lives, who can access it, which models process it and what those models are permitted to do.
Open models such as Apertus reduce dependence on closed APIs and let organisations choose their own infrastructure. But Swiss-developed ≠ Swiss-hosted, open source ≠ secure by default, and Swiss hosting alone does not establish legal compliance.
Own the documents, prompts, embeddings, logs and retention policy. Decide where each dataset is stored and processed.
Choose the Apertus version, review its license, evaluate the weights, update or fine-tune where suitable and avoid mandatory proprietary APIs.
Control the inference endpoint, hosting location, network boundaries, keys, backups and subcontractor dependencies.
Use your own identities, permissions, approval policies, monitoring, human oversight and change-management processes.
Retain exportable data, documented interfaces and the ability to change models or deployment providers without rebuilding your business.
“Swiss-made” on this page describes the Swiss research origin of the Apertus models. It is not a claim of an official Swiss Made certification, an automatic Swiss data-residency guarantee or endorsement of Pulsanto by ETH Zurich, EPFL or CSCS.
From on-premises inference to Swiss-based hosting, Pulsanto designs deployments around the sensitivity of your documents, your internal IT resources and your governance requirements.
We assess accuracy, latency, hardware availability, concurrency, data sensitivity and total cost. Quantisation, retrieval and targeted task design can often matter more than parameter count alone.
Apertus can be used through self-hosted model-serving frameworks and providers. The official project lists multiple Swiss-based options; specific service capabilities and contractual protections must be verified.
Use official Hugging Face checkpoints, supported inference libraries such as vLLM and your chosen hardware and network environment. You operate the service.
Select an approved provider offering Apertus inference under a reviewed service agreement, with clear logging, subprocessor, network and region controls.
Keep documents and retrieval under your control; send only explicitly approved context through a policy-governed model gateway.
Examples listed within the Apertus ecosystem in 2026, not Pulsanto endorsements. Offerings, availability, prices, regions and security terms may change.
A practical private AI platform combines a foundation model with your data layer, identity controls, retrieval pipeline and monitored integrations.
Contracts, documents, policies, records, approved audio and images
Classification, permissions, parsing, PII handling and data policies
Encrypted document storage, retrieval index, citations and context
On-prem or approved Swiss-hosted serving and controlled prompts
Assistants, MCP/API connections, audited actions and human approval
Data classification, minimal collection, TLS, encryption at rest, keys, backups, retention and deletion procedures.
SSO or corporate identity where appropriate, role-based and document-level access, tenant separation and MFA.
Prompt-injection testing, restricted tool scopes, approval gates, safe output handling and adversarial evaluation.
Audit trails, monitoring, patching, model versioning, incident response, vendor due diligence and tested recovery.
Applicable privacy and sector-specific obligations must be assessed per use case. Hosting location, an open model or encryption alone do not constitute a compliance certification. We design technical controls and evidence that support your organisation's legal and governance review.
Private AI becomes valuable when it solves a specific business process. Each application below can be designed with governed access to enterprise data — not an unrestricted public chatbot.
Ask questions across internal policies, product documentation, contracts and knowledge bases. Retrieve permission-aware answers with evidence and source references.
Extract obligations, compare clauses, track legal changes, build regulatory timelines and prepare drafts for qualified human review.
Classify incoming files, extract structured fields, reconcile information and flag inconsistencies, while preserving an auditable approval step.
Support multilingual search, correspondence and document workflows. Evaluate terminology and accuracy for German, French, Italian, English and relevant additional languages.
Summarise tickets, suggest responses, surface approved knowledge and route cases to the right person — without uncontrolled disclosure of internal records.
Use Apertus 1.5 to reason about text, images and audio inputs. Combine with separate OCR, transcription or business validation when the task requires it.
Connect to approved databases, calendars, ticketing tools or enterprise APIs with restricted scopes, audit logs and explicit approval of sensitive actions.
Build dedicated assistants for insurance, property, family administration, professional services and regulated knowledge workflows.
A download is not a deployment. Pulsanto combines AI engineering, enterprise software, secure document workflows and real-world use cases to make private LLMs useful in everyday operations.
Map business value, stakeholders, information sensitivity, residency requirements and risk constraints. Define evaluation criteria.
Compare Apertus versions, prompts, RAG, tool integrations, benchmark quality and assess performance on representative approved data.
Deploy inference, add identity and access management, encrypted storage, logging, monitoring, backup and approval policies.
Measure user adoption, evaluate failure cases, govern model updates, document ownership and build a sustainable operating process.
AI-powered household workflows, sensitive document management and structured personal knowledge in a Swiss platform.
View project ↗ 02 / LEGAL INTELLIGENCEHELVETIC LENSApertus-based regulatory intelligence organised around WATCH → DETECT → INTERPRET → ACT.
View demo ↗Locally deployed Apertus systems, document and contract processing, knowledge assistants and enterprise workflow integration.
Request demo ↗These examples describe our practical software and AI implementation experience. They do not imply an endorsement, affiliation or official technology partnership with ETH Zurich, EPFL, CSCS or the Apertus project.
Practical answers about the models, data protection and what it takes to build sovereign AI.
An open model gives access to model weights and supporting artefacts under its license. A private LLM is a deployment and governance choice: who runs the inference, who sees prompts, where documents and logs are stored, and which tools the model may invoke. Apertus supports private deployment; it does not configure that privacy on its own.
Yes. Apertus is developed within the Swiss AI Initiative by research teams associated with ETH Zurich, EPFL and CSCS, and trained on the Alps supercomputer in Lugano. This describes the model's origin — it does not certify where a particular business deployment is hosted.
The models are published under Apache 2.0, which permits commercial use under its terms. Deployment still entails GPU/CPU capacity, infrastructure, engineering, operations and security costs. Some model repositories require acceptance of the relevant usage policy and access conditions.
Apertus 1.5 provides 8B and 70B variants, text and image understanding, experimental audio understanding, text generation, improved instruction following and tool use, optional thinking mode and support for context lengths up to 262,144 tokens, subject to configuration. The earlier Apertus 1.0 was text-only; Mini models focus on smaller deployments.
Yes, with compatible model weights, serving software and adequate hardware. An isolated deployment can avoid external inference requests, but the engineering team must also inspect dependencies, telemetry, backups, update workflows and all application-level network calls. Isolation is a system property, not a single model setting.
The right choice depends on accuracy targets, multilingual requirements, tool reliability, concurrency, latency, hardware and operating cost. 8B is more accessible for local or constrained deployments; 70B generally needs much more compute. We benchmark both on actual business tasks before deciding. Quantisation and retrieval design also affect results.
Not when the full processing path is deliberately designed, configured and verified to remain in Switzerland. But a Swiss-developed model may be deployed anywhere. Storage, inference, logs, monitoring, remote support, subcontractors and disaster-recovery locations all need to be covered by your architecture and contractual controls.
Using a self-managed Apertus inference service does not inherently send your prompts to the model's researchers for training. However, application operators can choose to keep logs or process data in external services. Retention, telemetry, provider terms and any fine-tuning pipeline must be inspected and controlled.
No. Compliance depends on the specific application, its purposes, sensitive-data flows, lawful grounds, agreements, retention, access controls, transparency and applicable law. We implement privacy- and security-oriented technical controls to support governance and professional legal review, not automatic certification.
Apertus has broad multilingual capabilities; the original family was trained using more than 1,000 languages. That does not mean equivalent quality in all languages or tasks. For German, French, Italian, Romansh or mixed-language use cases, build a representative evaluation set and validate specialist terminology.
Usually not as a first step. Retrieval-augmented generation (RAG) can connect the model to indexed, permission-aware documents without modifying model weights. Fine-tuning becomes useful when repeated evaluation shows a need for task-specific behaviour and suitable training data is available.
Yes, through a controlled tool layer such as APIs and MCP servers. Each connector needs its own permissions, authentication, logging and limits. High-impact actions such as sending messages, changing records or initiating payments should require appropriate confirmation or human review.
Combine RAG with high-quality sources, citations, domain evaluations, constrained outputs, confidence or abstention rules, human approval for sensitive decisions and monitoring of production failures. No language model, including Apertus, guarantees factual accuracy on every response.
No such affiliation is claimed. Apertus is developed by the Swiss research institutions. Pulsanto independently implements AI applications and has practical experience with Apertus-based and private enterprise AI workflows.
Model availability and technical capabilities evolve. Use the primary project documentation and repositories for release-specific instructions, licensing and deployment requirements.
Information reviewed: October 2026Let's assess your use case, information sensitivity, model requirements and deployment options — then build a credible path to production.