DE · 2 Models
European providers
German model developer known for the Pharia-1-LLM family and the PhariaAI stack. Weights of the 7B control models are published on Hugging Face under a research-only licence, while commercial use runs through an on-premise agreement.
German company (Aleph Alpha GmbH, Heidelberg) subject to EU and German law; it markets "sovereign" on-premise deployments to European public-sector and industrial customers.
PT · 3 Models
European providers
European research project building open language models for all EU languages plus a set of additional languages. Weights and model cards are published openly under the utter-project organisation.
EU-funded research consortium led from Instituto Superior Técnico in Lisbon; the partners are European universities and companies, and the work is funded by the European Union rather than by a single commercial parent.
DE · 1 Models
European providers
The company behind this catalogue. LLM EU operates the router and the EU hosting layer, and has announced a first-party model, LLMEU V2, that is not built yet. Its own entries are listed here under the same evidence rules as every other provider: what is not published is marked unverified.
LLM EU is a private company, not an EU body, agency or institution. It publishes a catalogue, routes inference under policy, and intends to host first-party models on European infrastructure. Nothing in this catalogue implies an official status or an assessment by a public authority.
FR · 10 Models
European providers
French model developer that publishes both open-weight and commercially licensed models. Its line-up covers general instruct models, small edge models, a code model and audio models.
French company (Mistral AI, Paris), so the controller is bound by EU law and by the French implementation of the GDPR.
CZ · 2 Models
European providers
EU-funded project that plans to release a series of openly documented foundation models for European languages, including training data documentation and evaluation code. As of September 2026 no model weights have been released.
Consortium of European universities, HPC centres and companies, funded by the EU Digital Europe Programme under grant agreement 101195233; there is no single commercial parent company.
DE · 1 Models
European providers
German publicly funded project that trained the Teuken-7B model family on all 24 official EU languages. The released checkpoints are research licences rather than permissive open-source licences.
German research consortium (Fraunhofer, Forschungszentrum Jülich, TU Dresden, DFKI) funded by the German Federal Ministry for Economic Affairs and Climate Action, so the work falls under German and EU law.
CN · 8 Models
Chinese developer of the Yi family of bilingual Chinese-English language models, including the Yi-1.5 generations and the Yi-Coder code models.
Chinese company (Beijing), founded by Kai-Fu Lee. The Yi weights are published under Apache-2.0.
CN · 41 Models
Alibaba’s Qwen team publishes one of the broadest open-weight families, covering dense and mixture-of-experts text models, vision-language models, coder models and retrieval components.
Chinese company subject to Chinese law, including data and security regulations that apply to the parent group; the checkpoints listed in this catalogue are published under Apache-2.0.
US · 15 Models
US non-profit research institute behind the OLMo family of fully open language models, the Molmo vision-language models and the olmOCR document models. Releases include training code, data and evaluation harnesses.
US non-profit research institute in Seattle, Washington. OLMo weights, training code and data are released under Apache-2.0, which makes the release unusually auditable for a self-hosting operator.
CN · 5 Models
Chinese technology company publishing the ERNIE 4.5 family of open-weight mixture-of-experts language and vision-language models, plus OCR-specialised document models.
Chinese company (Beijing). The ERNIE 4.5 open weights are published under Apache-2.0; Baidu is not the controller of a self-hosted deployment in the EU.
US · 5 Models
Open research collaboration that trained the BLOOM family, a 176B-parameter multilingual model released under the custom Responsible AI Licence with a list of prohibited uses.
The BigScience workshop was coordinated by Hugging Face (US company) with hundreds of volunteer researchers; the BLOOM weights were trained on the French Jean Zay supercomputer. The Responsible AI Licence is a custom licence with use restrictions, not an OSI licence.
CA · 6 Models
Canadian lab whose Command line targets retrieval-augmented generation and tool use, and whose Aya line covers 23 languages with a European-language emphasis.
Canadian company subject to Canadian law; the open Command and Aya checkpoints use CC-BY-NC-4.0, which excludes commercial serving, while the newer Command A+ release is Apache-2.0.
CN · 21 Models
Chinese lab that publishes large mixture-of-experts models and reasoning models under the MIT licence, including distilled variants small enough to run on a single server.
Chinese company (Hangzhou) subject to Chinese law; its published weights use the MIT licence, but the hosted API and the company itself are outside EU jurisdiction.
US · 12 Models
US research collective known for the GPT-Neo, GPT-NeoX and Pythia families. Pythia is a suite of small models with 154 checkpoints published specifically so that training dynamics can be studied.
US research collective. The Pythia and GPT-Neo/X weights are published under Apache-2.0 and MIT respectively; EleutherAI is a research group, not a commercial vendor with a support contract.
US · 17 Models
Publisher of the Gemma open-weight family, including multimodal and audio-capable small models and an embedding model for retrieval.
US parent company subject to US law; Gemma weights ship under Google’s own Gemma terms rather than an OSI licence, and the Gemma 4 generation is published under Apache-2.0.
US · 17 Models
US technology company publishing the Granite family of small language models, embeddings, vision-language models and document-conversion models. Granite weights are released under Apache-2.0 with a published training-data description.
US company (IBM Corporation, Armonk, New York). The Granite weights are published under Apache-2.0, so a European operator runs them under its own responsibility; IBM is not the controller of a self-hosted deployment.
US · 15 Models
Publisher of the Llama family, the most widely deployed open-weight model line. Weights are released under the Llama community licences, which permit commercial use below a monthly-active-user threshold and add acceptable-use terms.
US company subject to US law; the Llama community licences are governed by California law and add use restrictions, and Meta platforms outside the EEA are the contracting entity for licensees outside the EEA and Switzerland.
US · 18 Models
Publisher of the Phi small-model family, including a multimodal variant with speech input, and of the multilingual-e5 text embedding model.
US company subject to US law; the Phi models are released under MIT, while Microsoft’s EU data-boundary commitments apply to its cloud services, not to the self-hosted weights.
CN · 6 Models
Chinese developer of the MiniMax text and multimodal families, including the M-series sparse mixture-of-experts models and the MiniMax-Text hybrid-attention models.
Chinese company (Shanghai). The MiniMax model licence is a custom agreement, not an OSI licence; it permits commercial use subject to its own use restrictions.
CN · 10 Models
Chinese developer of the Kimi models. The K2 and K3 series are very large sparse mixture-of-experts checkpoints released as open weights under a modified MIT licence.
Chinese company (Beijing). Kimi weights are published under a modified MIT licence that adds an attribution requirement to the user interface for larger deployments.
US · 18 Models
Publishes the Nemotron family of open-weight models and rerankers, aimed at long-context retrieval and agentic workloads, together with the evaluation harnesses used to score them.
US company subject to US law; the Nemotron weights carry the NVIDIA Nemotron Open Model License, a vendor licence rather than an OSI-approved one.
US · 11 Models
US lab that released the gpt-oss open-weight reasoning models and the Whisper speech-recognition family. Those two lines are the parts of its catalogue that can be self-hosted.
US company subject to US law; only some models are published as open weights, and those carry the Apache-2.0 or MIT licence independent of OpenAI’s own terms of use.
US · 326 Models
A broker, not an operator: requests are routed to whichever upstream provider OpenRouter selects, mostly outside the EU.
A US broker that routes to many model providers. Not an EU operator and not an EU region.
US · 4 Models
US enterprise software company whose research division published BLIP-2 and InstructBLIP, vision-language models that pair a frozen image encoder with a language model.
US company (San Francisco). BLIP-2 weights are MIT-licensed; the InstructBLIP Vicuna variant inherits a non-commercial base model and is carded as such.
CN · 6 Models
Chinese research laboratory publishing the InternLM and Intern-S1 model families, including chat models and multimodal scientific models.
Chinese state-backed research laboratory in Shanghai. InternLM weights carry custom licence terms with attribution requirements rather than a plain OSI licence.
GB · 3 Models
UK company publishing the StableLM family of small language models; the earlier checkpoints use the custom StableLM community licence.
UK company (London). The United Kingdom is a third country for EU data-protection purposes, and the StableLM community licence is a custom licence with a revenue threshold rather than an OSI licence.
AE · 10 Models
Research institute of the Abu Dhabi government, publisher of the Falcon family: dense transformers, Mamba hybrids and the Falcon-H1 hybrid SSM models. Licences are custom Falcon terms with an acceptable-use policy.
Abu Dhabi government research institute in the United Arab Emirates. The Falcon and Falcon-H1 licences are custom terms rather than an OSI licence, and the acceptable-use policy travels with the weights.
CN · 8 Models
Chinese technology company publishing the Hunyuan family: mixture-of-experts chat models, OCR and document models, and dedicated machine-translation models. The community licence excludes the EU from its territory.
Chinese company (Shenzhen). The Tencent Hunyuan community licences are territory-limited: their terms exclude the European Union, the United Kingdom and South Korea, which is why a European operator cannot rely on them.
CN · 16 Models
Chinese lab publishing the GLM family, including very large mixture-of-experts models, small flash models and a dedicated OCR model.
Chinese company subject to Chinese law; the GLM weights have moved between MIT and bespoke Z.ai licences, so each checkpoint has to be checked individually before commercial serving.