Do you want to call a model, download one, or run a platform?
Three different commitments, and the answer determines which half of this category you should read.
Call a model over an API: Mistral's La Plateforme, from roughly €1 per million input tokens for Mistral Small, processed in France. Hugging Face's Inference API serves open models the same way.
Download and self-host: Mistral publishes open weights including Mistral Large 3, Magistral Small for reasoning and the compact Ministral 3 family at 3B and 7B. Silo AI's Poro and Viking are Apache 2.0. Hugging Face is where all of it is distributed.
Run a governed platform: Aleph Alpha's PhariaAI and LightOn's Paradigm both deploy on-premise or in European private cloud with audit trails, role-based access and retrieval over your own documents. That is a procurement decision, not an API key.
How close is Europe to the frontier?
Close in specific places and behind at the very top, and both halves of that are worth saying.
Mistral Large 3 is a sparse mixture-of-experts model with 41 billion active parameters out of 675 billion total, reasoning across a 256,000-token context window, multimodal and multilingual from the start — and released as an open-weight download, which no US frontier lab matches at that tier. Magistral covers reasoning, Ministral 3 covers edge deployment at 3B and 7B, Voxtral covers audio.
The largest foundation models are still produced by US firms, and LightOn says as much in its own positioning: the scale of investment by OpenAI and Google is not currently matched in Europe.
Where Europe wins is narrower and real: DeepL on translation quality, Silo AI on Nordic languages, Aleph Alpha and LightOn on deployment inside a regulated boundary. Choosing a European model is a trade of maximum general capability for capability that fits the specific job.
What does a sovereign AI platform actually give you?
Data that never leaves, and an audit trail proving it — which is what turns an AI project from a pilot into something a regulator can be shown.
LightOn's Paradigm deploys on-premise or in European private cloud, serving LightOn's own models alongside open-source alternatives through one interface, with retrieval-augmented generation over proprietary documents and databases so nothing is sent outside. Its 2025 agentic functionality orchestrates tools across multi-step workflows — document analysis, extraction, summaries, comparisons across large document sets. Fine-tuned models stay inside the customer's infrastructure, so the training data and the resulting specialised model are never exposed.
Aleph Alpha's PhariaAI covers the same ground differently, running natively on STACKIT with hybrid execution that splits workloads by data classification, plus source attribution showing which documents produced a given answer — built against EU AI Act transparency requirements rather than retrofitted.
LightOn's reference clients make the point concrete: the Île-de-France region, Safran, Groupama and CNES, the French space agency. Those are organisations that cannot use a US API at all.
What happens to languages the big models neglect?
Somebody in Europe trains a model for them, and Silo AI is the clearest example of that working.
Poro is a 34-billion-parameter open model for English, Finnish and code, trained on 512 AMD MI250X GPUs on the LUMI supercomputer in Kajaani in collaboration with the University of Turku — genuinely understanding Finnish, a language US-centric models have always handled badly. Viking extends this to all the Nordic languages: Finnish, Swedish, Norwegian, Danish, Icelandic and English plus code, in 7B, 13B and 33B sizes, trained on two trillion tokens, outperforming general-purpose models like GPT on Nordic benchmarks.
Both are Apache 2.0, downloadable and deployable without licensing fees, and Silo AI published the open-source training framework so others can replicate the work. Viking is also available on Google Cloud for organisations preferring hosted deployment.
The broader point is that LUMI, funded by a consortium of European countries, made this possible — European infrastructure training European models on European data. Silo AI was acquired by AMD in 2024, which brings resources and raises a fair question about long-term commitment to Nordic open models; as of early 2026 that commitment appears intact.
Where does DeepL fit among language models?
As the counterexample to the assumption that European AI is always the compromise option.
DeepL is a neural machine translation system rather than a general-purpose LLM, and on the task it does — translation between 30-plus languages — it is widely judged better than Google Translate, particularly on European language pairs where nuance, register and idiom decide whether a translation is usable. That is the one place where the European product is not the trade-off.
Around it sit DeepL Write for writing assistance, document translation preserving formatting, a glossary feature enforcing consistent terminology across everything translated, and an API for embedding it in products.
DeepL is based in Cologne and processes under GDPR, which matters because translation input is routinely confidential — contracts, medical records, internal strategy pasted into a box by someone who has not thought about where it goes. Free tier, paid from €8.99 per month.
