Title: About LLM EU %E2%80%94 catalogue, router, host Canonical: https://llmeu.com/en/about/ Markdown mirror: https://llmeu.com/en/about.md Machine-readable: https://llmeu.com/llms.txt (index) - https://llmeu.com/llms-full.txt (full catalogue) - https://llmeu.com/llms-catalogue.json # About LLM EU Europe does not need another chatbot. It needs the place where its models live. ## Why this exists Most AI products in Europe are a chat window with somebody else's model behind it. The model itself was trained somewhere else, runs somewhere else, and the facts a buyer needs before choosing it %E2%80%94 who operates it, where a request is processed, which licence applies, which version is pinned %E2%80%94 live in scattered documents that are read after the purchase, if at all. LLM EU is built on the opposite assumption: what Europe needs is not another assistant but the place where its models can be documented, compared and run. That place has to be checkable. A data protection officer, a procurement team and a journalist should be able to read the same page and get the same answer, without asking us first. ## What the product is Three things, shipped as one platform. - **A catalogue** %E2%80%94 Model cards, providers, hosting regions, licences, benchmarks, task guides and the documents behind them. Where LLM EU cannot cite a source for a field, the card marks it unverified instead of filling it in. A model no endpoint serves is labelled catalogue only %E2%80%94 not hosted on LLM EU yet. - **A router** %E2%80%94 One OpenAI-compatible API. A request carries a policy; the router filters the endpoints that policy permits, scores what survives and takes the highest score. When nothing satisfies the policy, the call fails with 409 no_endpoint_for_policy. It is not rerouted somewhere else to look successful, and failover cannot step outside the filtered set. - **A host** %E2%80%94 First-party inference on European infrastructure: the region and the sovereignty class are recorded on the endpoint row rather than asserted in copy, and every response carries a llmeu block naming the model actually used, the region, the sovereignty class, the policy and the retention that applied. In this deployment the European rows are planned rather than serving, and the next section states that in numbers. [The catalogue](/en/models/) %C2%B7[Pricing per endpoint](/en/pricing/) %C2%B7[The API documentation](/en/docs/) %C2%B7 [The migration guide](/en/docs/migration/) ## Where requests go today LLM EU is EU-first: operated from Europe, built around European residency, and honest about the state of a given deployment on the day you read it. These are the endpoint rows this page rendered from. 325 endpoints are live, and none is in an EU region: every live row is in Outside the EU (OpenRouter broker) (non_eu_third_country), outside the EU. No live row records zero retention: the platform does not store prompt or completion bodies, and what the upstream provider retains is that provider's own claim, recorded on the row and not upgraded by us. 297 further endpoints name an EU region and are recorded as planned: they are the European capacity being built. Until they are live, a request that names an EU residency fails with 409 no_endpoint_for_policy %E2%80%94 the filter working, not an outage, and not a reason to relax it. Every count in this section is counted from the endpoint rows at render time, and one row per live endpoint is published on the residency map. [Every live endpoint, one row each](/en/studies/eu-llm-residency-map/) %C2%B7 [Residency on the model cards](/en/data-residency/) ## What "verified" means here A verified claim is one that names its source: a document you can open, or a value recorded on the endpoint row. A tick says the field is documented. It does not say that an auditor checked it, that the provider is correct, or that it will still be true next month. Verified is a documentation status. It is not a certificate, not an audit, not an AI Act conformity assessment, and not a statement about the provider's own compliance. LLM EU is a private company; it has no authority to certify anything and does not claim one. [What a verified card means](/en/verified/methodology/) %C2%B7 [How a claim is reviewed](/en/methodology/) ## How a fact gets onto a page A field is written down only after its source has been opened: provider documentation, a licence text, a weights repository or the endpoint row itself. The source travels with the value. Where the source does not answer the question, the field is published as unverified and named in the card's missing items, and an empty field is information rather than a gap to be guessed at. A verified card also records when it was last reviewed. The review is performed by LLM EU, not by a third party, and that is the limit of what the date means: it dates our reading of a document, not the provider's behaviour. [The review process, step by step](/en/methodology/) ## Zero retention on the platform, and what that does not cover LLM EU does not write prompt or completion bodies to its database. A trace records the model, the endpoint, the region, the token counts, the latency and the cost %E2%80%94 not the text. Only zero retention is implemented: a policy may offer 24h, 7d or 30d choices, but storing a body is not built, so a request whose effective retention would not be zero fails with 501 retention_not_implemented, and one that asks for more than its policy allows fails with 403 retention_exceeds_policy. The refusal is the feature. That is a property of this platform, not a statement about an upstream provider. Where an endpoint belongs to a partner, what that provider retains after the response is its own claim, recorded on the endpoint row: a row that does not record zero retention is shown as not zero, never assumed to be zero because our own database keeps nothing. [What is recorded, and what is not](/en/docs/evidence/) %C2%B7[The privacy notice](/en/legal/privacy/) %C2%B7 [Retention on the model cards](/en/verified/retention/) ## What LLM EU does not claim These are not modesty. They are the claims the product cannot support, so they are refused in public rather than implied. - No certification, no third-party audit, no conformity assessment. None exists, and none is claimed. - No EU institution, no supervisory authority, no official status. LLM EU is a private company. - No claim that inference runs in the EU today. The EU-hosted endpoint rows are planned; the live ones are not in an EU region, and the section above states how many of each. - No claim that a provider retains nothing. The platform keeps no prompt bodies; a provider's retention is its own claim, and a row that does not record zero retention is published as not zero. - No claim that a catalogue row is a running deployment. A model with no endpoint is catalogue only %E2%80%94 not hosted on LLM EU yet. - No invented numbers. A parameter count, context window or benchmark score with no source is published as unverified, and a field nobody documented is left empty. - No customer logo, testimonial or case study appears on this site. If that changes, the reference will be published with the permission behind it. - No legal entity details beyond what the imprint publishes. Until those details are configured by the operator, the imprint says so on its own page instead of naming a company that has not been entered. ## Who operates this site LLM EU is a private company product. The company name, registered address and register details belong on the imprint, and that page is the only place this site publishes them. The imprint is generated from the deployment's own configuration (LEGAL_ENTITY_NAME, LEGAL_ENTITY_ADDRESS, LEGAL_ENTITY_EMAIL, and optionally register, VAT ID, legal form, director and phone). While those are unset, the page renders as a draft, states in its own words that the entity details are missing, and is deliberately not indexed. A half-filled imprint is never presented to a crawler as a finished legal document. [The imprint](/en/legal/imprint/) ## Contact One address reaches the operator of this site: mail@llmeu.com. There is no contact form here %E2%80%94 the public site is served as static files, so there is nothing that could receive a submission %E2%80%94 and no support phone number is published. [What to include in a message](/en/contact/) ## Related - [The model catalogue](/en/models/) - [Pricing](/en/pricing/) - [Documentation](/en/docs/) - [Legal documents](/en/legal/) - [EU LLM price index](/en/studies/eu-llm-price-index/) ## Read the catalogue or call the API The catalogue is public and needs no account. The API is OpenAI-compatible, and a new account starts with the free credit described on the pricing page. [Read the catalogue or call the API](/en/free-llm-api/)