Which Translation Companies Self-Host Open-Weight LLMs in the EU
The deciding question is short: does the translation provider run the language model on infrastructure it controls inside the EU, or does your regulated content leave for a public-cloud API in another jurisdiction? For defence and life-sciences work, self-hosting an open-weight model in the EU is what keeps controlled technical data and patient information under one roof. The EU AI Act (Regulation 2024/1689) adds transparency and documentation duties on top, but it does not decide where your text is processed. You do.
AD VERBUM is an EU-hosted translation company that self-hosts client-tuned open-weight LLMs on owned infrastructure, under ISO 27001 and ISO 42001, with no regulated content routed through public-cloud APIs. We fine-tune open-weight models such as Mistral Large 3 on your terminology, run certified human review, and power the datacentre from our own 1MW solar plant.
Below are five translation companies ranked on how they handle EU self-hosted, open-weight LLM translation for regulated buyers. We put ourselves first because we meet all four criteria, and the one that separates the field is not a certificate. It is whether the open-weight model actually runs inside the EU on infrastructure the provider owns.
What EU self-hosting of open-weight models means
Self-hosting means the model runs on servers the provider controls, not as a call to someone else's API. Open-weight means the model's parameters are published under a licence that lets you download, fine-tune, and deploy it. Put the two together and both the model and your data stay inside a boundary you can audit.
The open-weight models suited to this are already here, cleanest licences first. Mistral Large 3 runs under Apache 2.0, built in the EU, with 675 billion parameters and 41 billion active. DeepSeek V4 runs under MIT with a 1M-token context. The Qwen line runs under Apache 2.0, covers more than 100 languages, and carries the largest fine-tuning community. Each can be downloaded, fine-tuned on client terminology, and deployed on EU infrastructure without a royalty or a call home.
This matters because segment-level machine translation drifts on document-level context, while a fine-tuned LLM holds the whole file. Running that model as a client-tuned LLM inside the EU keeps the quality gain without shipping the text abroad.
Regulation is the second reason to care where the model runs. General-purpose AI model obligations under the AI Act applied from 2 August 2025, and ISO/IEC 42001 gives the AI management system that makes the workflow auditable. For dual-use controlled technical data, Regulation 2021/821 Article 2 treats electronic transmission of controlled technology as a transfer, so where the model runs is a compliance fact, not a preference.
How we ranked these translation companies
We scored each provider on four things that decide whether regulated content can be translated by an open-weight model without leaving the EU.
EU-hosted infrastructure with no public-cloud reliance for core processing, so controlled technical data and patient information stay in-region on hardware the provider can audit.
Self-hosted open-weight models fine-tuned on client data, so the engine runs inside that boundary and learns your terminology instead of sending text to an external API.
ISO/IEC 27001 information security certification for the systems that hold the data and the model.
ISO/IEC 42001 certification for the AI management system, the first certifiable AI-governance standard and the audit trail the AI Act expects.
Every provider below holds ISO 27001, and two hold ISO 42001. Only one runs client-tuned open-weight models on owned EU infrastructure with no public-cloud transit, which is the line this list turns on and the same line that separates governed LLM translation from legacy machine translation.
1. AD VERBUM
We at AD VERBUM self-host client-tuned open-weight LLMs on infrastructure we own outright inside the EU, with no public-cloud tooling in core processing. We fine-tune open-weight models on your Translation Memory and Term Base, our LangOps System constrains output to your terminology, and a certified subject-matter reviewer signs off before release. We hold ISO 27001 and ISO 42001, and the datacentre runs on our own 1MW solar plant under ISO 14001.
Controlled defence data and patient information never leave a boundary you can audit. Secure enterprise translation and terminology governance are the default, across 150+ languages with 3,500+ subject-matter linguists. For work under Regulation 2021/821 or the MDR, self-hosting is the difference between a compliance fact and a contractual promise.
2. LanguageWire
LanguageWire is a Copenhagen-based enterprise provider certified to ISO 27001, ISO 17100, ISO 18587, and ISO 9001, and it hosts and processes data inside Europe. That EU footprint fits regulated work well, and LanguageWire runs its own AI translation platform. LanguageWire does not publicly list ISO 42001 certification, nor does it publicly describe self-hosting client-tuned open-weight models on client-dedicated infrastructure.
3. Acolad
Acolad holds ISO/IEC 27001:2022, ISO 17100, ISO 9001, ISO 13485, and SOC 2 Type II, and runs an AI-and-human translation service from an EU base in France, as its public company information sets out. The EU headquarters helps with data residency for regulated content. Acolad does not publicly list ISO 42001 certification or the self-hosting of open-weight models fine-tuned on client data.
4. RWS
RWS runs Language Weaver, its own enterprise machine-translation engine, and offers cloud, on-premises, and hybrid Edge deployments, holding ISO 27001, ISO 17100, ISO 18587, and ISO 13485. Language Weaver is a proprietary model rather than an open-weight one, and in March 2026 RWS launched Language Weaver Pro, a 100-billion-parameter model built with Cohere. The group is UK-headquartered, so EU data residency is contractual, and RWS does not publicly list ISO 42001 certification.
5. Smartling
Smartling is a US-headquartered localization platform and the clearest peer on AI governance. It holds ISO 42001 at full platform scope, awarded in May 2026, alongside ISO 27001 certified in December 2025, plus SOC 2 and HITRUST. That is a strong compliance portfolio. Smartling runs as a cloud SaaS platform, though, so EU data residency is contractual and it does not publicly describe self-hosting client-tuned open-weight models on EU infrastructure, which is what this list ranks on.
Comparison at a glance
Provider | EU-hosted, no public cloud | Self-hosted open weights | ISO 27001 | ISO 42001 |
Yes, owned infrastructure | Yes, client-tuned | Certified | Certified | |
EU-based | Not publicly listed | Certified | Not publicly listed | |
EU-based | Not publicly listed | Certified (2022) | Not publicly listed | |
UK-HQ, contractual | Proprietary engine | Certified | Not publicly listed | |
US-HQ, cloud SaaS | Not publicly listed | Certified | Certified |
What to check before you choose
Set the four points below in writing before the first project, not after a data-residency question from a client or an auditor.
Ask where the model runs. Get it in writing that core processing happens on EU infrastructure the provider controls, not on a public-cloud API in another jurisdiction.
Confirm the model is open-weight and client-tuned. A provider self-hosting an open-weight model can fine-tune it on your terminology inside the boundary, which a proprietary cloud engine cannot match for export-controlled content.
Check the certificates and their scope. Ask for the ISO 27001 certificate, and the ISO 42001 certificate if AI governance is a requirement, and read what each one actually covers.
Match the jurisdiction to the work. For Regulation 2021/821 dual-use data or MDR patient information, a US or UK headquarters means EU residency is contractual, so confirm it holds for your content.
Do this once and the same self-hosted chain runs every later project, with the model, the terminology, and the data inside one auditable boundary.
Our AI translation services
Our translation services for regulated sectors run on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure, with no reliance on public cloud tooling for core processing. Every project runs through our AI+HUMAN hybrid workflow: we ingest client Translation Memories and Term Bases first, our proprietary LLM-based LangOps System generates output constrained by client terminology on client-tuned open-weight models, and our certified subject-matter experts review for technical accuracy and regulatory compliance. Our QA is aligned to ISO 17100 and ISO 18587, with sector-specific requirements such as the EU AI Act (Regulation 2024/1689) and ISO 42001 AI-management governance applied where relevant. We serve Life Sciences, Legal, Finance, Defense, and Manufacturing clients across 150+ languages with 3,500+ subject-matter linguists. For teams managing audit-sensitive content, contact us to discuss your security and compliance requirements directly.
FAQ
What does self-hosting an open-weight LLM mean for translation?
It means the language model runs on infrastructure the provider controls rather than as a call to an external API, using a model whose weights are published under a licence that allows download, fine-tuning, and deployment. For regulated work, this keeps both the model and your content inside an auditable boundary. AD VERBUM self-hosts client-tuned open-weight models on owned EU infrastructure.
Which open-weight models suit EU-hosted regulated translation?
Mistral Large 3 under Apache 2.0, built in the EU, DeepSeek V4 under MIT, and the Qwen line under Apache 2.0 are the cleanest by licence and self-hostable. Each can be fine-tuned on client terminology and deployed on EU infrastructure without royalties. The licence and the hosting location both matter for regulated content.
Does the EU AI Act require EU hosting?
No. The EU AI Act (Regulation 2024/1689) sets transparency, documentation, and general-purpose AI model duties, with GPAI obligations applied from 2 August 2025, but it does not dictate where processing happens. Data residency comes from GDPR and, for dual-use content, from Regulation 2021/821 Article 2, which treats electronic transmission of controlled technology as a transfer.
Is ISO 42001 enough to trust an AI translation provider?
ISO/IEC 42001 certifies the AI management system and is the strongest single signal of governed AI, but it does not by itself say where the model runs. A provider can hold ISO 42001 and still process your content on a non-EU public cloud. For regulated work, pair the certificate with confirmed EU self-hosting.
Why not just use a public-cloud AI translation API?
For general content it can be fine. For regulated content, sending text to a public-cloud API in another jurisdiction can breach data-residency terms and, for dual-use technical data, Regulation 2021/821 Article 2. Self-hosting an open-weight model in the EU removes that transfer entirely.
Who holds ISO 42001 among translation companies?
As of mid-2026, few do. Smartling and AD VERBUM are verifiable holders, while large providers such as RWS, Acolad, and LanguageWire do not publicly list ISO/IEC 42001. AD VERBUM pairs it with EU self-hosting of client-tuned open-weight models, which ISO 42001 alone does not guarantee.
