Regulatory update — June 2026
This article refers to an August 2026 deadline for the EU AI Act's high-risk (Annex III) obligations. Under the EU's Digital Omnibus — provisionally agreed in 2026 but not yet final law — that deadline is now deferred to 2 December 2027 (high-risk AI embedded in regulated products under Annex I is deferred further still). Treat the August 2026 dates referenced below as the original statutory timeline, and confirm the current status before relying on either date.
In September 2025, Switzerland released Apertus — a fully open, multilingual large language model trained on the Alps supercomputer in Lugano. Six weeks later, Basel publisher Helbing Lichtenhahn announced Swiss-Noxtua, a sovereign legal AI platform integrating the Basler Kommentar and Commentaire romand. And in February 2026, Noxtua launched Europe's first cross-border legal AI license, covering Germany, Austria, and Switzerland in a single subscription.
These are not incremental improvements. They represent a structural shift: for the first time, DACH law firms can choose AI tools built around European infrastructure and jurisdiction-specific legal content, making EU AI Act and Swiss professional-secrecy analysis materially more tractable than with generic US-hosted tools.
This article analyses what these developments mean in practice — which models exist, what they can and cannot do, how to evaluate them, and what the regulatory landscape demands before August 2026.
Swiss legal tech companies
19 funded, 9 at Series A or later. Switzerland has become Europe's densest legal tech ecosystem relative to its size.
The Swiss LLM Landscape: What Actually Exists
Apertus — Switzerland's National AI
Apertus is the flagship output of the Swiss National AI Initiative, a joint project of ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). It is not a legal tool — it is a general-purpose foundation model designed to be fine-tuned for specific domains.
Key specifications:
- Parameters: 8B and 70B variants
- Training data: 15 trillion tokens across 1,000+ languages, 40% non-English
- Architecture: Decoder-only transformer with novel xIELU activation, AdEMAMix optimiser, QRPO alignment
- Languages: All Swiss national languages (German, French, Italian, Romansh) plus English and 1,000+ others
- License: Apache 2.0 — fully open for commercial use, fine-tuning, and redistribution
- Infrastructure: Trained on 10,000+ NVIDIA Grace Hopper Superchips, 100% renewable energy
- EU AI Act: First large model designed for full EU AI Act compliance from inception
What it does well: Multilingual text processing, translation between Swiss languages (German-Romansh BLEU 27.8 vs Llama-3.3's 21.6), long-context tasks (80.6% on benchmarks).
What it does not do well: Apertus-70B scores 69.6% on MMLU versus Claude 3.5 Sonnet's 88.7% and Llama 3.1 70B's 83.6%. Math and coding performance trail frontier models by 15-20 points. Swiss German dialectal output is described by reviewers as "practically unusable." The team explicitly states this first iteration prioritises "safety and accessibility over building a frontier model."
Apertus Is Not a Frontier Model
Do not evaluate Apertus against GPT-4 or Claude on general benchmarks. That is not the comparison that matters for legal practice. The relevant comparison is: can you fine-tune it on your firm's legal corpus, deploy it on Swiss infrastructure, and maintain full control over client data — without needing permission from a US provider? On that axis, no closed-source model competes.
The Fine-Tuning Path: From General to Legal
Apertus's Apache 2.0 license means any firm can fine-tune it on proprietary legal data. The Swiss AI Initiative has published official fine-tuning recipes on GitHub (swiss-ai/apertus-finetuning-recipes) supporting both LoRA and full fine-tuning via standard frameworks (TRL, Accelerate, Transformers).
A legal-specific fine-tuning project exists at swiss-ai/apertus-legal, though as of April 2026 it contains compliance documentation rather than a fine-tuned model. The initiative has announced domain-specific adaptations for law, health, climate, and education — but the legal version has not been released yet.
What fine-tuning requires in practice:
- Compute: The 70B model requires significant GPU resources for fine-tuning. LoRA reduces this to a single high-VRAM GPU (48-80GB). The 8B model can be LoRA fine-tuned on a single RTX 4090 or A100.
- Data: A curated corpus of Swiss/German/Austrian legal texts — legislation, case law, commentary, contracts. Quality and jurisdiction-specificity matter more than volume.
- Evaluation: Legal domain benchmarks do not exist for Swiss law. Firms fine-tuning Apertus will need to build their own evaluation sets — a non-trivial investment.
- Hosting: Swisscom Swiss AI Platform offers native Apertus access with sovereign Swiss hosting. Safe Swiss Cloud and Infomaniak provide additional options.
The performance gap between models has narrowed — curated, jurisdiction-specific legal content is now the differentiator, not model size.
Noxtua: The DACH Legal AI Platform
While Apertus is a foundation model waiting to be specialised, Noxtua is already purpose-built for legal work. Understanding the distinction is critical for any firm evaluating its options.
What Noxtua is: A sovereign legal AI platform built on a proprietary Legal LLM trained on 55+ million legal documents from Europe's leading legal publishers. It is not based on Apertus — it uses its own model architecture and a proprietary search model (Noxtua Voyage Embed).
The publisher alliance: Noxtua's competitive advantage is its exclusive partnerships with the dominant legal publishers in each DACH market:
- Germany: C.H. Beck — Beck-Noxtua
- Austria: MANZ — MANZ-Noxtua (launched October 2025)
- Switzerland: Helbing Lichtenhahn (Basler Kommentar, Commentaire romand, legalis.ch) — Swiss-Noxtua
This means Noxtua's AI is grounded in the authoritative commentary that DACH lawyers already rely on — not scraped web data or general corpora.
Infrastructure: All data processed on European servers (IONOS and Open Telekom Cloud). No US data flow. Swiss-Noxtua specifically operates from Swiss data centres.
Capabilities: Legal research, analysis ("Understanding"), and drafting — in German, French, Italian, and English.
Europe License: Launched February 2026, this is the first cross-border legal AI license. A single subscription covers multiple European jurisdictions — significant for firms doing cross-border work across DACH.
Funding: EUR 80.7 million Series B (April 2025), led by C.H. Beck, with CMS, Northern Data, and Dentons participating. The team quadrupled to approximately 80 people within one year.
€80.7M
Noxtua Series B
Led by C.H. Beck, with CMS, Dentons, Northern Data
$41.2M
DeepJudge Series A
Felicis + Coatue; 3 ex-Google/ETH Zurich founders
$200M
Harvey AI raise
March 2026 financing round at a reported $11B valuation
The Broader Swiss Legal Tech Ecosystem
Noxtua and DeepJudge are the largest players, but Switzerland's legal tech ecosystem includes 93 companies:
DeepJudge (Zurich) — USD 41.2 million Series A, founded by three ex-Google researchers with ETH Zurich PhDs. AI-powered enterprise search that turns a firm's internal knowledge into searchable intelligence. Clients include Freshfields, CMS Switzerland, Schoenherr, and Wenger Vieli. Not a model producer — a platform that uses existing LLMs for legal workflows.
Legartis — Contract analytics and intelligence. Specialises in German, English, and French document analysis with jurisdiction-specific models for Swiss, German, and Austrian law.
CASUS / Contractus.AI — Contract drafting, review, and risk detection. Word integration plus web app — practical tools for daily contract work.
BALO.AI — AI-powered anonymisation of court decisions and sensitive legal documents. Addresses a specific DACH need: Swiss and German courts publish decisions with redacted personal data, and the anonymisation process is labour-intensive.
Unplex — AI-driven legal automation with on-premise deployment — directly addressing the FADP and BGFA compliance requirements that make cloud deployment problematic for some firms.
Why Data Sovereignty Is Not Optional for DACH Law Firms
The argument for sovereign AI in legal practice is not philosophical — it is structural. Three overlapping legal obligations create hard constraints that US-hosted AI tools cannot satisfy by default:
1. Swiss Professional Secrecy (Art. 321 StGB)
Breach of professional secrecy carries up to three years imprisonment and may also trigger a monetary penalty under the general Swiss criminal-law framework. The offence is subject to the ordinary Swiss criminal-law limitation rules. A firm using a US-hosted AI tool to draft client documents may be in breach before any data protection regulation enters the picture. The Swiss Bar Association (SAV) identifies three compliant pathways: internal deployment, compliant outsourcing where the provider qualifies as a Hilfsperson, or informed client consent — which the Federal Court has questioned can genuinely be "informed" given AI opacity.
2. The US CLOUD Act Conflict
The US CLOUD Act empowers US authorities to compel US-based providers to produce data stored anywhere in the world. That creates a structural conflict with Swiss and EU secrecy and data-protection obligations. Customer-managed encryption with keys held outside provider infrastructure is one of the few meaningful mitigations, but it does not remove the need to review Swiss-law secrecy, subcontracting chains, and operational access rights vendor by vendor.
3. Swiss FADP and EU AI Act
The Swiss FADP is directly applicable to AI systems processing personal data — the Federal Data Protection and Information Commissioner confirmed this in May 2025. For law firms, the FADP imposes personal criminal liability on managing directors (not companies) for intentional violations, with penalties up to CHF 250,000. The EU AI Act, fully applicable from August 2026, classifies AI used in the "administration of justice" as high-risk — triggering conformity assessments, technical documentation, and human oversight requirements.
The Dual Compliance Trap
Swiss firms face a unique regulatory position: they must comply with the Swiss FADP and professional secrecy rules (Art. 321 StGB) for domestic operations, AND the EU AI Act for any work involving EU clients or EU-market AI tools — which covers virtually every major legal AI platform. Compliance with one framework does not guarantee compliance with the other.
Three Regulatory Regimes, Three Countries
The regulatory landscape across DACH is not uniform. Understanding the differences matters for cross-border firms:
Switzerland has chosen a sector-specific approach to AI regulation. The Federal Council decided in February 2025 not to create a comprehensive AI Act. Instead, Switzerland will ratify the Council of Europe's Framework Convention on AI and integrate AI governance into existing sectoral laws. A bill for consultation is expected by end of 2026. This creates a window of flexibility — but also uncertainty about which sector-specific rules will apply to legal AI.
Germany applies the EU AI Act directly. The German Federal Bar Association (Bundesrechtsanwaltskammer) is developing comprehensive AI ethics opinions focused on transparency requirements and professional responsibility. German law firms using AI for litigation support or regulatory advice face high-risk classification under Annex III of the EU AI Act.
Austria applies the EU AI Act directly. For DACH firms, the practical issue is less a separate Austrian AI statute than how Austrian regulators, courts, and clients interpret documentation, sovereignty, and deployment expectations under the EU framework.
Sep 2023
Swiss FADP enters force
New data protection act directly applicable to AI. Personal criminal liability for managing directors up to CHF 250,000.
In forceAug 2024
EU AI Act entered into force
Prohibited practices and AI literacy obligations already applicable since February 2025. GPAI governance since August 2025.
Phasing inFeb 2025
Swiss sector-specific approach decided
Federal Council chooses not to create a comprehensive AI Act. Sector-specific amendments and Council of Europe Convention ratification planned.
Policy setAug 2026
EU AI Act full applicability
High-risk provisions take effect. Legal AI tools used in administration of justice trigger conformity assessments, documentation, and human oversight.
Deadline
Apertus vs Noxtua vs US Models: A Decision Framework
The choice is not binary. Different use cases call for different tools. Here is how to evaluate them:
Use Apertus when:
- You need full control over model weights and training data
- Your firm has the technical capacity for fine-tuning (or can engage a Swiss AI services provider)
- You want to build proprietary legal AI capabilities that no competitor can access
- On-premise deployment is a hard requirement
- Budget for fine-tuning and evaluation exists
Use Noxtua when:
- You need a production-ready legal research and drafting tool now, not in six months
- Your work spans DACH jurisdictions and you need cross-border legal content
- Integration with authoritative commentary (Basler Kommentar, beck-online, MANZ) matters
- You prefer a managed platform over building and maintaining your own
Use Harvey or other US-based tools when:
- Your work is primarily anglophone or international (not DACH-specific)
- The specific capabilities of the tool outweigh the data sovereignty trade-off
- Your firm has secured adequate data processing agreements and client consent
- Professional secrecy obligations do not apply (non-privileged work only)
Use a combination when:
- Most firms will. Sovereign tools for privileged client work. International tools for non-sensitive research. This is not all-or-nothing.
Which AI Strategy Fits Your Firm?
5 questions
1. Where is most of your client data currently processed?
2. Does your firm handle cross-border DACH work (CH/DE/AT)?
3. What is your firm's technical capacity for AI deployment?
4. How do you handle professional secrecy compliance for digital tools?
5. What is your primary use case for legal AI?
Hosting Options: Where to Run Legal AI in Switzerland
For firms that need to control where their AI processes data, five Swiss and European options exist:
Scroll horizontally to view all columns.
What Is Coming: The Next Twelve Months
Apertus legal fine-tuning — The Swiss AI Initiative has announced domain-specific adaptations for law, based on the Apertus foundation. A legal-specific model trained on Swiss legal corpora would be the first open-source, multilingual legal LLM built from Swiss infrastructure. Timeline is not public, but the compliance documentation groundwork exists.
Noxtua expansion — The Europe License (February 2026) is Noxtua's bridge to scale beyond DACH. Expect additional publisher partnerships and jurisdiction coverage. The Neur.on acquisition (Fribourg NLP team) strengthens Swiss quadrilingual capabilities.
EU AI Act enforcement — August 2, 2026 is the hard deadline for high-risk AI provisions. Firms using AI for litigation support, regulatory advice, or HR decisions must have conformity assessments, technical documentation, and human oversight mechanisms in place.
Legal AI agents — Both DeepJudge and Harvey launched agentic AI products in early 2026. Noxtua identifies agents as one of three trends shaping European legal AI. These are AI systems that execute multi-step legal workflows autonomously — a qualitative shift from the current chat-and-search paradigm.
Swiss AI regulation consultation — The Federal Department of Justice and Police (FDJP) will present a bill for consultation by end of 2026. This will clarify which sector-specific rules apply to legal AI in Switzerland.
Swiss LLM Evaluation Checklist for Law Firms
0/10Key Takeaway
The choice between Swiss/European and US-based AI is not primarily about model quality — frontier models still outperform Apertus on general benchmarks. It is about data sovereignty, professional secrecy compliance, and jurisdiction-specific legal content. For privileged client work in DACH, sovereign AI is not a preference — it is a professional obligation. Start with managed platforms (Noxtua) for immediate productivity, evaluate Apertus fine-tuning for long-term competitive advantage, and use international tools only for non-privileged work with adequate safeguards.
Art. 321
Swiss professional secrecy
Criminal-law analysis applies independently of data-protection analysis.
Aug 2026
EU AI Act deadline
High-risk provisions fully applicable
Feb 2026
Europe License launched
Noxtua introduced a cross-border license spanning multiple jurisdictions.
Adriana Adafinoaiei advises law firms on AI adoption strategy, data sovereignty compliance, and EU AI Act readiness across Switzerland, Germany, and Austria. To discuss your firm's AI strategy, get in touch.