European AI model labs

Aleph Alpha trains language models in Heidelberg. Its latest release, Kolibri, became available on 3 October 2026. It handles German and English text, document questions and tool calls for public administration and industry. The company provides downloadable model weights and licence terms.

On 16 September 2026, Aleph Alpha and Cohere signed a combination agreement. The proposed group would use the Cohere name and have headquarters in Berlin and Toronto. Aleph Alpha's October release still said regulatory approval was pending and it continued to operate independently. Its last standalone funding round was more than $500 million on 6 November 2023.

Domyn has its headquarters in Milan. The company, previously called iGenius, trains Domyn Large and Domyn Small for enterprise reasoning and language tasks. Large has proprietary access; Small has open weights. On 20 October 2026 it began a staged release of Domyn Edge, a four-billion-parameter model for tool use and multi-step tasks, with training checkpoints, recipes and code. It says Edge was trained from scratch on CINECA's Leonardo supercomputer.

Founder Uljan Sharka told Sifted that Domyn had raised $1.1 billion in debt and equity over two years, reported on 20 August 2026. That figure is combined financing across the period, rather than the size of one equity round.

H Company is a French company registered in Paris that trains computer-use models. Its Holo models read screens and produce actions for browser, desktop and mobile tasks. Releases include Holo1, Holo1.5, Holo2, Holo3, Holo3.1 and Holo4. Holo4 35B-A3B and Holo3.1 have Apache 2.0 weights; Holo4 27B has research-only weights. Holo1.5 and Holo2 weights use CC BY-NC 4.0. The catalogue also includes Holotron-12B and Holotron 3 Nano under NVIDIA’s Open Model License, and Apache 2.0 NeoMME encoders and retrievers for document understanding.

Charles Kantor remains a shareholder. Other investors include Accel, UiPath, Bpifrance, Amazon and Elaia, which participated in the $220 million seed round announced on 21 May 2024. H is a member of NVIDIA’s Nemotron Coalition.

LightOn SA is registered in Paris and trains its own models. It co-developed ModernBERT with Answer.AI for text encoding, classification and retrieval. Its open-weight releases include GTE-ModernColBERT for document retrieval, LightOnOCR, LightOnOCR-2 and LightOnOCR-3 for reading documents and extracting layout, and DenseOn, LateOn, their multilingual versions, SlideOn and mSlideOn for search. Earlier commercial releases include Alfred for text generation. Its model catalogue also includes colgrep-agent-2B for searching code.

LightOn is listed on Euronext Growth Paris. Its four co-founders together owned 2,868,742 shares on 23 September 2026. That day it announced bond financing of up to €4 million nominal, with €3.7 million in subscription proceeds; €1.84 million was paid immediately and the remaining subscription depended on conditions. Orange Business and Hewlett Packard Enterprise are commercial partners.

Mistral has its headquarters in Paris and trains language, coding and speech models. Its Mistral Large 4 preview, announced on 6 October 2026, targets reasoning, coding and tool-based work. Large 4 weights were released on 29 October under the Mistral Ecosystem Open-Weight License, which restricts some model-as-a-service uses. The multimodal model accepts text and images and generates text. The lab also introduced Voxtral speech models in July 2025.

Mistral announced a €3 billion Series D at a valuation above €21 billion, led by Samsung Electronics. Its news index dates the announcement 8 September 2026.

Almawave develops the Velvet family of language models. Its 2024 results presentation describes the foundation models as trained from scratch and documents a multilingual corpus with a substantial Italian share. The models support text applications for businesses and public bodies. Almawave Labs S.r.l. names itself as the developer responsible for Velvet training and gives its registered office in Rome. Almawave is part of the Almaviva group.

Corti develops models for clinical speech, documentation and reasoning. The company says its Symphony model was trained from scratch on patient data. Its APIs let software developers add transcription and clinical documentation to healthcare applications. Corti ApS has its principal office in Copenhagen and is incorporated in Denmark. Corti also runs open-weight models from other developers through a separate hosting service.

DeepL SE has its legal seat in Cologne, Germany, and trains its own language models. Its releases include the next-generation translation LLM, the proprietary LLM behind DeepL Write Pro, and Voice models for live speech translation. These are supplied through DeepL’s products and APIs; model weights are not offered in those releases.

Its investors include Index Ventures, ICONIQ Growth, Teachers’ Venture Growth, IVP, Atomico and WiL. On 22 May 2024, DeepL disclosed $300 million in investment at a $2 billion valuation. Deutsche Bahn, Nikkei, Coursera and Zendesk are named customers.

Engineering develops EngGPT 2, the language model inside its IS-IA software platform. The company says it trained the 16-billion-parameter model from scratch on 2.5 trillion tokens, with roughly a quarter of the data in Italian. The platform adds tools for adapting models and connecting them to business processes. Engineering Ingegneria Informatica S.p.A. has its registered office in Rome.

Gladia was founded in Paris and acquired by OVH Groupe on 31 July 2026. It trains the Solaria speech-recognition family: Solaria-3, released on 10 June 2026, targets conversational recordings in English, French, German, Spanish and Italian. Solaria-1 covers a broader range of languages. Both are offered through the API.

Its last venture round was a $16 million Series A in October 2024. The company says it will retain its own brand and customer API within OVH Groupe.

Gradium builds speech generation, transcription and live translation models from its Paris base. It also releases Phonon for speech generation on devices: the 31-million-parameter English model has an Apache 2.0 licence, while the 40-million-parameter variants have non-commercial licences with separate commercial terms. On 8 July 2026, Gradium extended its seed financing to $100 million and announced a San Francisco Bay Area office.

Neuphonic is a London speech model company. It releases NeuTTS-2M for multilingual speech and voice cloning, NeuTTS-2E for English speech with emotion control, NeuCodec for audio compression and NeuDecide for speech-to-tool calls. The models can run on customer hardware; downloadable weights have different licences. NeuCodec and NeuDecide use Apache 2.0, while NeuTTS-2E has commercial-use conditions under the NeuTTS Open License.

Its July 2026 NeuTTS-2E release has a roughly 125-million-parameter active backbone and four fixed speakers. It released NeuDecide in October 2026. Its latest disclosed funding is €3.5 million in pre-seed financing announced on 17 October 2024, led by Moonfire VC.

Phonexia develops speech recognition, speaker identification and voice-analysis models. Its 2022 product release explicitly describes a Mandarin speech-to-text model trained from scratch. Its current documentation also describes self-supervised training for voice deepfake detection. The company supplies software that customers can integrate into transcription, security and call-analysis systems. Phonexia s.r.o. has its registered seat in Brno, Czechia.

Silo AI is a Helsinki AI company acquired by AMD. Its SiloGen team developed Poro-34B with the University of Turku and the HPLT project. Poro was pretrained on Finnish, English and code, with weights released under Apache 2.0. The collaboration also produced the Nordic Viking models. Silo AI remains a Finnish legal company within the US-owned AMD group; its later Poro 2 work uses continued pretraining of existing models.

Speechmatics is headquartered in Cambridge and trains speech-recognition models. Its Oak medical model, released on 14 October 2026, transcribes consultations across languages and speakers while tracking clinical terminology. Oak is available through the API, with cloud, private-cloud and on-premise deployment. The release identifies no open-weight distribution.

Its latest disclosed funding is its $62 million Series B on 28 June 2022, led by Susquehanna Growth Equity.

Tilde develops language models and translation and speech software from Riga. Its TildeOpen-30B is a multilingual base language model trained on the LUMI supercomputer. The model card documents its tokenizer, training data and pretraining schedule, with particular attention to smaller European languages. Tilde publishes the weights under CC BY 4.0. The base release needs further training to follow instructions or work as a chat assistant.

Black Forest Labs trains the FLUX family for image generation and editing. Its 2026 FLUX 3 offering also includes video with audio and action prediction for robotics. The lab is based in Freiburg im Breisgau.

Its FLUX 3 Image Klein release on 15 October 2026 has 5.6 billion parameters and open weights. Other FLUX models are available through the company's API, so access terms depend on the variant. Its latest funding round was a $300 million Series B in December 2025 at a $3.25 billion valuation, co-led by Salesforce Ventures and Anjney Midha.

Mirelo trains audio models in Berlin. Its Mirelo SFX 2.0 generates effects and ambience around the action in a video, with an option to preserve dialogue. It became generally available on 14 October 2026 through Studio, the API and editor integrations. These are hosted releases; the cited pages identify no open-weight download.

Mirelo announced a $41 million seed round on 15 December 2025, co-led by Andreessen Horowitz and Index Ventures.

Basecamp Research trains EDEN, a 28-billion-parameter biological model, on evolutionary sequence data. From its London base, it uses the model to design proteins, peptides and gene-editing tools. Researchers can use two EDEN capabilities through a Claude connector for antibiotic design and immune-response prediction. The model page offers those tools and papers without a public weight download.

On 23 September 2026 it announced a $140 million Series C led by S32. The funding supports further EDEN training and the development of its own therapeutic candidates toward clinical trials.

BeyondMath develops a physics foundation model in London for simulating airflow, aerodynamics and heat transfer. Its Generative Physics offering lets engineers upload geometry and simulate designs.

On 25 February 2026 it announced a $10 million seed extension, led by Cambridge Innovation Capital. That brought the seed round to $18.5 million. The funding supports research and commercial deployment.

Bioptimus is a Paris biology model lab. Its H-Optimus family learns representations of tissue-slide images for pathology research. H-Optimus-0 has Apache 2.0 weights; H-Optimus-1 and H0-mini have downloadable weights under non-commercial research licences. Its M-Optimus model combines pathology, spatial transcriptomics and genomics and has a commercial deployment licence.

Bioptimus disclosed a $41 million investment led by Cathay Innovation on 14 January 2025, taking total funding to $76 million. In March 2026 it launched STELA, a programme to generate linked spatial biology data for M-Optimus.

Cradle has its headquarters in Amsterdam and a Swiss legal entity, Cradle Switzerland AG, in Zurich. It trains protein models using public datasets, its Amsterdam wet-lab data and customers’ assay results. Its commercial releases are protein-design and optimization models supplied through a web application and API. Cradle does not give these models public release names. They remain private within customer workspaces, with access sold through annual software licences.

Its investors include IVP, Index Ventures and Kindred Capital. It announced a $73 million Series B led by IVP on 26 November 2024. Customers named in that announcement include Novo Nordisk, Johnson & Johnson Innovative Medicine, Novonesis and Grifols. UCB announced a collaboration on 4 November 2026 to use Cradle in antibody discovery and engineering.

CuspAI is headquartered in Cambridge and incorporated in England and Wales as Cusp AI Limited. It trains materials foundation models using crystal structures, properties, text and images. Its deployed system generates candidate materials, evaluates them with physics simulations and updates its models from experiments. It supplies the models through industrial partnerships, without a named public checkpoint or open-weight release.

Investors include Kleiner Perkins, NEA and Bezos Expeditions. On 20 July 2026, CuspAI announced a $450 million Series B at a $2.6 billion valuation. Its AI Materials Foundry partners include NVIDIA, Meta, Samsung, Hyundai Motor Group, Henkel and Applied Materials.

Jua trains EPT-2 weather models in Zurich. They forecast wind, temperature and solar radiation for energy trading and renewable power operations. The company distributes forecasts through an API; its public model offering does not identify downloadable weights. Its 2026 releases include longer-range EPT-2e forecasts and the EPT-2.1 Helios solar model. Jua announced an $11 million Series A on 4 June 2025, co-led by Ananda Impact Ventures and Future Energy Ventures.

Latent Labs builds its own generative protein models from bases in London and San Francisco. Latent-X2, announced in December 2025, designs antibodies; its Latent-Y drug-design agent opened to researchers in July 2026. Access is through the hosted platform. It disclosed $50 million in total funding, including a $40 million Series A, on 13 February 2025.

Neuralk-AI is a Paris company founded in 2023. It trains a tabular foundation model for classification and regression on business data. In April 2026 it renamed the model Seldon; an older company timeline also calls it Dawn. Access is through an API, and these pages identify no downloadable model weights. It disclosed a $4 million seed round in February 2025, led by Fly Ventures.

NXAI develops recurrent neural networks and pretrained models in Linz. Its xLSTM release includes a seven-billion-parameter language model trained on 2.3 trillion tokens. The company also publishes Vision-LSTM image models and TiRex, a pretrained model for time-series forecasting. TiRex can make forecasts on new datasets without a separate training run for each dataset. The repositories provide weights, code and links to the research papers.

PhysicsX is based in London and trains models that predict how physical designs behave. Its PX-Car-Aero-v1, introduced on 7 August 2026, predicts drag, lift, pressure and airflow from vehicle geometry. PhysicsX trained it on more than 33,000 simulated designs and makes it available through its platform for customers to adapt to their vehicles. The release describes platform access and does not announce open weights.

The company raised a $300 million Series C at an approximately $2.4 billion valuation on 8 June 2026. Temasek led the round, with M&G Investments, Intrepid Growth Partners and existing investors participating.

Prior Labs trains TabPFN models for predictions from tables, including classification and regression. It keeps its headquarters in Freiburg. Its current TabPFN-3.5 offering addresses structured, grouped and temporal data. Weights and deployment licences are available under specified terms, including a commercial licence; an open software package does not make every weight release unrestricted.

SAP completed the acquisition in July 2026. Prior Labs continues as a separate entity inside SAP, which committed more than €1 billion over four years to expand the lab. SAP's October 2026 quarterly statement reports approximately €518 million in preliminary acquisition consideration. That purchase price and the future investment commitment are separate amounts.

SpAItial builds Echo world models for 3D environments from London. Echo-2 generates worlds from text or images; Echo-3, released in October 2026, reconstructs spaces from phone captures. Access is through its apps and API. It announced a $13 million seed round on 26 May 2025.

Wayve is based in London and its UK company is Wayve Technologies Ltd. It trains the AI Driver for vehicle control, LINGO-1 and LINGO-2 for driving commentary and language-conditioned driving, and the GAIA world-model family for simulation. GAIA-4, announced on 3 August 2026, generates camera and radar inputs as the AI Driver acts in a simulated scene. These research announcements demonstrate the models; they do not offer downloadable weights.

SoftBank Group holds more than 25% and no more than 50% of shares and voting rights. Wayve announced a $1.2 billion Series D at an $8.6 billion valuation on 25 February 2026, followed by a $60 million extension from AMD, Arm and Qualcomm on 15 April. It announced an $85 million employee share-sale offer on 1 July. Named partners include Nissan, Uber and Stellantis.

AI Sweden, RISE and WASP WARA Media & Language developed GPT-SW3 for Nordic language applications. The pretrained models have 126M, 356M, 1.3B, 6.7B, 20B and 40B parameters. The open-weight release followed a research preview in November 2023, under a modified Apache 2 licence described in the team’s paper. Access is now restricted to approved researchers at European universities and research institutes under the repository’s licence. Lindholmen Science Park hosts AI Sweden, which receives funding from Vinnova and its partners.

BSC’s Language Modeling team trains multilingual models for text generation, translation, summarisation and question answering. Its Salamandra family has 2B, 7B and 40B parameter models; ALIA-40B was pretrained from scratch and has instruction-following variants. Both families publish weights under Apache 2.0. The newer Salamandra-7b-fc-2610-evo adds reinforcement learning for instruction following and tool use, with access subject to the repository’s conditions. Spain’s Ministry for Digital Transformation funds the work through EU NextGenerationEU, with support from Catalonia’s Aina project. BSC is a public consortium of the Spanish government, Catalan government and Universitat Politècnica de Catalunya.

EuroLLM is developed by Instituto Superior Técnico, Instituto de Telecomunicações, Unbabel and university and industry partners across Europe. The team pretrained and released 1.7B, 9B and 22B models with Apache 2.0 weights for multilingual text tasks, with base and instruction-following versions. The latest release, EuroLLM-22B, arrived on 14 December 2025 with Apache 2.0 weights and support for 35 languages, including all 24 official EU languages. Uses include translation, summarisation and question answering. Funding includes Horizon Europe’s UTTER project, the ERC DECOLLAGE grant and Portugal’s Recovery and Resilience Plan; EuroHPC supplied training compute.

Kyutai is a French endowment fund headquartered in Paris. The nonprofit lab was founded by iliad, CMA CGM and Schmidt Futures, which announced nearly €300 million in funding on 17 November 2023. Its current donors are iliad, CMA CGM and Schmidt Sciences.

Its model releases include Moshi for spoken dialogue, Mimi for audio coding, Hibiki and Hibiki-Zero for speech translation, streaming speech-to-text and text-to-speech models, Pocket TTS for speech generation, Helium 1 for language, and MoshiVis for visual conversation. Downloadable weights accompany its open-model programme. Other releases include CASA for visual understanding, REKI for robot actions, and CADRE video autoencoders with noncommercial weights. MuScriptor transcribes music into MIDI. Kyutai also released the MIRA multiplayer world-model demo with General Intuition and Epic Games, and works with Les Invincibles on the Invincible Voice speech tool. These collaborations span gaming and assistive communication.

NorwAI is hosted by NTNU’s Department of Computer Science and funded by the Research Council of Norway and its partners. Its NorLLM programme trained NorGPT-369M, NorGPT-3B, NorGPT-23B and NorLlama-3B from scratch, then released Mistral-7B, Llama2-7B and Mixtral-8x7B adaptations for Norwegian text processing. Its latest model family addition is NorwAI-Magistral-24B, a reasoning model released in 2025. Weights require approval under the NorLLM licence, which restricts use geographically and bars resale. Applications include summarisation, document comprehension and text generation.

OpenEuroLLM brings together 20 research organisations, companies and EuroHPC centres, coordinated by Charles University and co-led by AMD Silo AI. Its latest main experimental release is Prelude, a 9.095B parameter base model covering 36 target languages, with public Apache 2.0 checkpoints available since July 2026. It supports research into pretraining and evaluation. Earlier joint releases include 38 HPLT monolingual 2.15B models and open-sci-ref 0.01 models at 130M, 400M, 1.3B and 1.7B, with public weights. The EU’s Digital Europe Programme funds the consortium under grant 101195233.

The SpeakLeash Foundation and ACK Cyfronet AGH develop Bielik for Polish text generation and instruction following. Released weights include Bielik-11B v3 and the smaller 1.5B and 4.5B v3 models, all under Apache 2.0. The small models continue pretraining from Qwen2.5; their actual parameter counts are 1.6B and 4.6B. PLGrid compute grants supported training on Cyfronet’s Athena and Helios supercomputers. The latest releases include DFlash draft models, whose weights were uploaded on 14 June 2026, for speculative decoding alongside Bielik target models.

ETH Zurich and EPFL co-lead the Swiss AI Initiative, working with CSCS on Apertus. The team pretrained and released Apertus 1.0 in 8B and 70B sizes in September 2025, with Apache 2.0 weights for multilingual applications and further adaptation. The latest release, Apertus 1.5, arrived in July 2026 in 8B and 70B sizes with Apache 2.0 weights with image and audio understanding, tool use and reasoning capabilities. Funding comes from the ETH Board, CSCS compute on Alps and strategic partners including Swisscom.