Every European LLM reviewed
Paris, France
Founded 2023
Free and open source models
Open weights
Best for: Anyone who needs real-time voice AI they can run and inspect
Kyutai is a non-profit research lab rather than a company, funded with €300 million from French backers including Xavier Niel and Rodolphe Saadé, and set up to do open science rather than to sell a product. That structure is unusual enough to matter: there is no commercial pressure to keep the good models closed.
Its flagship result is Moshi, a voice model that speaks and listens in real time rather than transcribing, thinking and then speaking. The theoretical latency is around 160 milliseconds and in practice 200 to 240 — which is the range where a conversation stops feeling like a query and starts feeling like a conversation. It was built from scratch by a small team in about six months, and released openly.
That release is the point. Real-time voice AI from the American labs is available as an API and nothing else; Kyutai publishes the model so it can be run locally, inspected, fine-tuned and built on without a per-minute bill or a network round trip.
The lab operates from Paris. What it is not is a product: there is no interface, no support and no service-level agreement, and using this means engineering. Its focus is voice rather than general-purpose language, so for a chat assistant Mistral remains the answer.
What Kyutai does well
- Moshi speaks and listens in real time at 200ms latency
- Models released openly rather than behind an API
- Non-profit lab with no pressure to close the good work
- Run locally with no per-minute cost or round trip
- Paris-based, funded at serious scale
Where Kyutai falls short
- A research lab, not a product — no interface or support
- Voice-focused rather than general-purpose language
- Using it means real engineering work
- No service-level agreement of any kind
Standout feature. Real-time conversational voice you can download — which every American lab offers only as a metered API.
Paris, France
Founded 2024
HoloTab Chrome extension free; Holo Models API self-serve via hub.hcompany.ai; Autonomous Enterprise Operations platform and custom model access priced through direct sales engagement.
Free HoloTab extension; self-serve Holo Models API signup
Best for: Enterprises automating browser and desktop tasks with AI agents
H Company is a Paris AI lab built around a simple bet: the next leap in AI is agents that act, not just answer.
Its Holo family of multimodal models is trained to interpret screens and interfaces the way a person would, then carry out multi-step tasks — filling forms, navigating sites, completing workflows — rather than describing how to do them. HoloTab, a free Chrome extension, puts that directly into a browser sidebar with a "Routines" feature for recurring chores. The company is young, founded in 2024, but backed by investors including Samsung, AWS, Bpifrance, Eurazeo and Xavier Niel's Iliad.
Beyond the free extension, the Holo Models API opens the same multimodal agent models to developers through a self-serve hub, and July 2026 brought Computer-use Agents — fully managed agents that take actions on a computer rather than just a browser tab. The Autonomous Enterprise Operations platform wraps models, agents, scalable execution and continuous learning into one offering aimed at enterprises that want AI doing real operational work, sold through a direct sales conversation rather than a price list.
The honest caveats: H Company publishes no public pricing beyond the free HoloTab extension, so evaluating the API or enterprise platform means a sales call. Its models are proprietary, with no open-weight release the way Mistral or Pleias offer.
And it is competing directly with the computer-use agents OpenAI and Anthropic are shipping from a much larger research and compute base, so track record matters more than the pitch. For an EU-jurisdiction alternative, though, it is doing the work rather than just positioning around it.
What H Company does well
- Free HoloTab browser agent to try immediately
- Holo Models API accessible via a self-serve hub
- Computer-use agents for full desktop automation
- Backed by Samsung, AWS, Bpifrance and Eurazeo
- Paris-based, contracts under French and EU law
Where H Company falls short
- No public pricing beyond the free extension
- Proprietary models, no open-weight release
- Founded 2024 — thin track record so far
- Competing directly with OpenAI and Anthropic on agents
Standout feature. A browser and computer-use agent that acts on tasks directly, not a chat window that only describes how.
Paris, France
Founded 2023
Common Corpus dataset and Pleias small language models (Pleias-RAG-1B, Pleias-Pico and others) free and open on Hugging Face under Apache 2.0; Synth and Stratum on-premise deployments priced through a demo call, no published rate card.
Free open models and Common Corpus dataset on Hugging Face; book a demo for Synth/Stratum
Best for: Teams needing open, rights-cleared training data and small models
Pleias occupies a different corner of the European AI map: instead of chasing frontier scale, it builds the data layer underneath — and increasingly, the small models trained on it.
Common Corpus, its flagship dataset, is described as the largest open, rights-cleared collection for LLM pre-training, built from government records, legal archives and scientific literature with clear provenance rather than scraped copyrighted text. The work won an ICLR Oral, and partners include Nvidia, Mozilla, the Wikimedia Foundation and the AI Alliance.
On top of that data, Pleias trains and publishes small language models — including Pleias-RAG-1B and the Pleias-Pico family — openly on Hugging Face under Apache 2.0, small enough to run offline on hardware costing under €100.
Sillon, a 600-million-parameter model built for Paris transport operator RATP, is a concrete production example of a small, specialised model reported to outperform general models many times its size on its specific task. Synth and Stratum extend the same approach into synthetic training data and document processing, both deployable fully on-premise.
The trade-off is scale and polish. Pleias is not building a ChatGPT competitor — there is no general-purpose flagship chat product, no public pricing page, and engaging Synth or Stratum means booking a demo rather than reading a price list. Founded in 2023 as a small Paris SAS, it is also young and thinly staffed next to Mistral. What it offers instead is a genuinely open, auditable data supply chain and small models with a traceable training history.
What Pleias does well
- Common Corpus: largest open, rights-cleared LLM training dataset
- Small open-weight models on Hugging Face under Apache 2.0
- Production track record (RATP's Sillon model)
- On-premise deployment for Synth and Stratum
- Backed by Nvidia, Mozilla and the Wikimedia Foundation as partners
Where Pleias falls short
- No general-purpose flagship chat product
- No public pricing for Synth or Stratum
- Small, young company founded in 2023
- Best fit is data and training pipelines, not an end-user assistant
Standout feature. Training data with known, rights-cleared provenance — the opposite of "trust us" that most LLM pre-training asks for.
Riga, Latvia
Founded 1991
Machine translation, speech, AI assistant and TildeOpen LLM products priced per enterprise engagement, quoted on request; TildeOpen-30B released free and open under CC BY 4.0 on Hugging Face.
Contact sales for a platform trial; TildeOpen LLM free to download
Best for: Baltic and CEE organisations needing language AI for under-served languages
Tilde has been building language technology in Riga since 1991, long before "AI" meant large language models, and that history shows in the breadth of what it now ships: machine translation, speech-to-text and voice synthesis, multilingual AI assistants for customer support and public-sector workplaces, and AI data services for fine-tuning. Clients skew towards governments and EU institutions, which fits a company whose core claim is thirty-plus years of Baltic and Central European language coverage most vendors treat as an afterthought.
TildeOpen-30B is the newer piece: a 30-billion-parameter open multilingual LLM covering EU languages that dominant models handle poorly, released free under CC BY 4.0 on Hugging Face rather than kept behind an API.
It sits alongside Tilde's translation and speech products, giving an organisation the option of an open model to inspect and self-host or a managed enterprise service with an SLA. ISO 27001 certification and GDPR-first positioning target exactly the public-sector and regulated-industry buyers who ask hard questions about where language data goes.
What Tilde is not is a general frontier lab: there is no public pricing page, evaluating the enterprise platform means a sales conversation, and TildeOpen-30B is not aimed at competing with Mistral or GPT-class models on general reasoning — its value is coverage of Latvian, Lithuanian, Estonian and other languages the frontier labs undertrain. For an organisation whose real language need is Baltic or Central European rather than English, that focus is the entire point.
What Tilde does well
- 30+ years building European language technology
- TildeOpen-30B open and free under CC BY 4.0
- Deep coverage of Baltic and CEE languages others neglect
- ISO 27001 certified, GDPR-first for public-sector buyers
- Full stack: translation, speech, assistants and an open LLM
Where Tilde falls short
- No public pricing, enterprise sales cycle to evaluate
- TildeOpen-30B not built to compete on general reasoning
- Smaller scale than Mistral or the large translation vendors
- Best fit is narrower than a general-purpose assistant
Standout feature. Thirty-plus years of Baltic and Central European language data behind one open 30-billion-parameter model.
Paris, France
Founded 2023
Free tier / pay-per-use API
Free tier
Best for: Developers and businesses wanting frontier-class models via API or self-hosted
Mistral is the answer to whether Europe has a frontier AI lab.
Mistral Large 3, released in December 2025, 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 with multimodal and multilingual capability built in — and available as an open-weight download, which is the part no US frontier lab matches.
Mistral Medium 3 covers the cost-efficiency middle, Magistral Small and Medium handle reasoning with Small open-sourced, Ministral 3 provides dense 3B and 7B models for edge deployment, and Voxtral covers audio.
Two products sit on top. Le Chat is the consumer assistant on web, iOS and Android with deep research, native multilingual reasoning, image editing and a Memories feature, free with Pro at $14.99 per month. La Plateforme is the developer platform, with API usage from roughly €1 per million input tokens for Mistral Small and open-weight self-hosting where the only cost is compute — which beats per-token pricing at scale.
Mistral AI operates from Paris, so API calls stay in EU jurisdiction under French law rather than falling under the CLOUD Act. The considerations are maturity rather than capability: the release cadence is fast enough that staying current takes attention, Le Chat trails ChatGPT on plugins and third-party integrations, and enterprise support is thinner than Google's or Microsoft's.
What Mistral AI does well
- Frontier-class models released as open weights
- 256,000-token context, multimodal and multilingual
- Full range from 3B edge models to 675B total parameters
- API in EU jurisdiction under French law
- From about €1 per million input tokens for Small
Where Mistral AI falls short
- Le Chat trails ChatGPT on plugins and integrations
- Enterprise support thinner than Microsoft or Google
- Fast release cadence requires ongoing attention
- Not every model in the lineup is open-weight
Standout feature. Frontier-class weights you can actually download — the only lab in the world at that tier that lets you.
Cologne, Germany
Founded 2017
Free tier / from €8.99 per month
Free tier
Best for: Anyone translating between European languages where quality decides usability
DeepL is the exception in European AI: a product that beats its American counterpart on the task it exists to do rather than matching it while offering better jurisdiction. Its neural machine translation across more than 30 languages is consistently judged stronger than Google Translate on European pairs, and the reason is nuance — register, idiom and the difference between a translation that is intelligible and one a native speaker would actually write.
The surrounding features are the ones professional use requires. Document translation preserves formatting rather than returning text you have to rebuild. The glossary enforces consistent terminology across everything translated, which for a company with product names, legal terms or internal vocabulary is the difference between usable output and constant correction. DeepL Write extends the same models into writing assistance, and the API embeds translation into products and workflows.
DeepL operates from Cologne and processes under GDPR, which matters more than it appears: translation input is routinely confidential — contracts, medical records, internal strategy — pasted in by someone who never considered where it goes. Free tier for occasional use, paid plans from €8.99 per month, and API pricing for volume. It does one thing, and it is not a general-purpose language model.
What DeepL does well
- Translation quality generally ahead of Google Translate on European pairs
- Glossaries enforce consistent terminology across all output
- Document translation preserving original formatting
- DeepL Write for writing assistance on the same models
- German company, GDPR processing, free tier available
Where DeepL falls short
- Translation only — not a general-purpose language model
- Fewer languages than Google Translate overall
- Advanced features require a paid plan
- Not open source and not self-hostable
Standout feature. The one European AI product where choosing it costs you nothing in quality — the translation is simply better.
Paris, France (and New York)
Free tier / Pro from $9 per month / Enterprise
Free tier
Best for: Developers and researchers working with open models
Hugging Face is the infrastructure of open machine learning. The Model Hub hosts hundreds of thousands of models with documentation, evaluation and versioning; the Transformers library is the standard way to load and run them; Datasets covers training data; Spaces hosts interactive applications so a model can be tried before it is committed to; and the Inference API serves models without building deployment infrastructure.
For a European organisation the sovereignty argument runs through it in a specific way. Hugging Face is not itself an EU-jurisdiction vendor — despite French origins, Hugging Face Inc. is US-incorporated and subject to US jurisdiction, which is a genuine consideration for strict sovereignty requirements. What it enables is the route around that question entirely: a downloaded open model running on your own hardware has no jurisdiction at all.
The other honest caveat is quality variance. The scale of the Hub means many models are well documented and thoroughly evaluated while others are experimental or abandoned, and telling them apart takes expertise. Large-scale inference through the managed API is also not the cheapest option at volume. Free tier, Pro from $9 per month, Enterprise above that.
What Hugging Face does well
- The distribution point for open models, datasets and demos
- Transformers library is the standard way to run them
- Spaces let you evaluate a model before committing
- Downloading a model removes the jurisdiction question entirely
- Generous free tier, Pro from $9/month
Where Hugging Face falls short
- US-incorporated despite French origins
- Model quality on the Hub varies widely
- Managed inference not cheapest at scale
- Requires ML familiarity to use well
Standout feature. The route to sovereignty runs through it rather than from it: download the weights and the jurisdiction question disappears.
Heidelberg, Germany
Founded 2019
Enterprise pricing on request
Contact sales
Best for: Government and regulated enterprises needing AI inside a hard data boundary
Aleph Alpha shifted from selling access to its Luminous models to shipping PhariaAI, an enterprise operating system for generative AI, and that shift is what makes it relevant. PhariaAI covers the full lifecycle — model hosting, orchestration, retrieval-augmented generation, explainability and compliance — running natively on STACKIT so compute, models and application layer are all operated by European companies under European law.
Hybrid execution is its distinctive capability: workloads run across on-premise and cloud simultaneously, with sensitive data processed only on local servers while less sensitive tasks overflow to cloud capacity, allocated by classification rules the organisation defines. That converts an all-or-nothing decision into a policy one — patient records stay local while general summarisation uses cloud capacity.
Explainability closes the loop. PhariaAI shows which source documents contributed to an output through attention visualisation and source attribution, so an AI-informed decision can be traced and reviewed — exactly what the EU AI Act asks of high-risk systems, built in rather than retrofitted. Aleph Alpha GmbH operates from Heidelberg with enterprise-only pricing on request, proprietary models rather than open weights, and raw capability below the largest US foundation models.
What Aleph Alpha does well
- Sovereign stack — compute, models and application under EU law
- Hybrid execution splits workloads by data classification
- Source attribution for auditable AI decisions
- Built against EU AI Act requirements from the start
- On-premise deployment for classified workloads
Where Aleph Alpha falls short
- Enterprise pricing only, nothing published
- Proprietary — no open weights
- Raw capability below the largest US models
- Sales-led with no self-service entry
Standout feature. Hybrid execution by data classification: the sensitive workload stays on-premise while the rest uses cloud capacity, on your rules.
Helsinki, Finland
Open-source models free / enterprise engagements on request
Open models on Hugging Face
Best for: Nordic organisations needing models that genuinely handle their languages
Silo AI built the models the large labs were never going to build. Poro, a 34-billion-parameter open model for English, Finnish and code, was trained on 512 AMD MI250X GPUs on the LUMI supercomputer in Kajaani in collaboration with the University of Turku — producing genuine Finnish understanding in a field where US-centric models have always treated Finnish as an afterthought.
Viking extends the work across all the Nordic languages — Finnish, Swedish, Norwegian, Danish, Icelandic and English plus programming languages — in 7B, 13B and 33B variants trained on two trillion tokens, achieving state-of-the-art results on Nordic benchmarks and significantly outperforming general-purpose models like GPT on those tasks.
Both are Apache 2.0, downloadable and deployable without licensing fees, and Silo AI published its open-source training framework so others can build on it. Viking is also available on Google Cloud for hosted deployment.
Beyond models, Silo AI runs enterprise AI engagements, custom model development, consulting and MLOps, priced for medium to large organisations — smaller businesses will mostly benefit from the open models. Two honest limits: Poro and Viking are not designed to compete with the largest general-purpose models on English reasoning, and the AMD acquisition in 2024 raises a fair question about long-term commitment to Nordic open models, though as of early 2026 that commitment appears to hold.
What Silo AI does well
- Poro and Viking are Apache 2.0 and free to deploy
- Outperform general-purpose models on Nordic benchmarks
- Trained on LUMI, European supercomputing infrastructure
- Open-source training framework published for others
- Sizes from 7B to 34B to fit available hardware
Where Silo AI falls short
- Not competitive with the largest models on English reasoning
- Enterprise engagements priced for larger organisations
- AMD ownership raises questions about long-term direction
- Narrow language focus by design
Standout feature. Models that actually understand Finnish and Icelandic — trained on a European supercomputer because nobody else was going to build them.
Paris, France
Enterprise licence pricing on request
Proof-of-concept deployment
Best for: French and European enterprises deploying generative AI inside their own infrastructure
LightOn's Paradigm is a generative AI platform designed to run where the data already is — on-premise or in European private cloud — serving LightOn's own models alongside popular open-source alternatives through a single interface. Retrieval-augmented generation over unstructured data lets an organisation build knowledge applications on its own documents and databases without sending anything to an external server, which is the requirement that rules out US APIs entirely for some buyers.
Its 2025 agentic functionality is the more interesting half: agents that select and orchestrate tools across multi-step workflows — searching document bases, exploring the web securely, analysing documents, extracting information, producing summaries and running comparisons across large document sets. Fine-tuning within Paradigm customises base models on proprietary data for legal, medical or engineering domains, and the fine-tuned models stay inside the customer's infrastructure so neither the training data nor the resulting model is ever exposed.
LightOn SAS is a Paris spin-off from École Polytechnique research and an active contributor of models to Hugging Face including Reason-ModernColBERT, GTE-ModernColBERT and BioClinical ModernBERT. Its reference clients — the Île-de-France region, Safran, Groupama and CNES — are the kind that cannot use a US AI API. Pricing is licence-based rather than per-token, so costs are predictable; it is enterprise-only, engagement is sales-led, and the largest foundation models remain American.
What LightOn does well
- On-premise or EU private cloud deployment, data never leaves
- Agentic workflows across large document sets
- Fine-tuned models stay in your infrastructure
- Licence pricing is predictable, not per-token
- Reference clients including Safran, Groupama and CNES
Where LightOn falls short
- Enterprise-only, unsuitable for startups wanting a quick API
- Pricing on request, sales-led engagement
- Deployment more complex than a US cloud API
- Largest foundation models are still American
Standout feature. Licence pricing rather than per-token: the cost of an AI rollout is a number you can budget rather than one you discover.
