European LLMs in 2026: The Complete Map

More than two dozen European organisations have set out to build their own AI models since 2023. Three have stopped, one is merging into a Canadian company, one is a promise, not a product, and none cracks the top ten of the main overall leaderboard, where every entry on 6 October was American.

This is the full map of European large language models (LLMs) and the home of our review series. Its companion is our Chinese LLM map. Updated 6 October 2026, the day Mistral announced Large 4, a 1 trillion parameter model with weights due on 27 October.

Why This Map Follows the Money

Germany’s OpenGPT-X consortium did everything by the book. It delivered Teuken-7B, an LLM that runs in all 24 EU languages, then wrapped up on time in early 2025 when its federal funding ran out. The model is still downloadable, but nobody is training the next one.

That is how most European LLM projects end: not in failure, but in a grant that finishes. Two companies have released open weights above 100 billion parameters on commercial money, Mistral and the Franco-American Poolside. Everyone else runs on public grants, academic compute, or an exit. So this map is organised by who is paying, for how long, and what happens when the money stops.

The Full List of European LLMs in 2026

Licences matter here more than parameter counts. “Open” is doing a lot of unearned work in European AI communication, so the column says what the licence file actually permits. Projects that stopped are in a separate table below.

Not on this map: Russia, where Sber released the MIT-licensed 432B GigaChat 3.5, and New York-headquartered Hugging Face.

Who Stopped Training Their Own Models?

Three have stopped. The weights are still downloadable, but nobody is training a successor.

Three Different Bets Under One Word

The commercial bet. Mistral carries most of it. On 6 October it announced Large 4, a 1 trillion parameter model with 49 billion active, trained on 4,000 Nvidia Grace Blackwell chips in European data centres, with the weights promised for 27 October. Its shipped flagships are Large 3 (675B) and Small 4 (119B) under Apache 2.0, plus Medium 3.5, whose licence shuts out companies above $20M in monthly revenue. Poolside released the weights of Laguna M.1 (225B) and S 2.1 (118B), then reportedly licensed the factory that built them to Nvidia. JetBrains trained its 12B Mellum2 from scratch under Apache 2.0.

The national champions. The crowded lane, mostly built on somebody else’s base: Bulgaria’s BgGPT, Slovenia’s GaMS and Norway’s Borealis on Google’s Gemma, Estonia’s EstLLM on Meta’s Llama and Hungary’s Racka on Alibaba’s Qwen. Poland has two families of its own, and Latvia’s Tilde trained a 30B model for 34 languages on EU compute. Portugal shipped Amália for €5.5M by extending EuroLLM, and Norway’s NorwAI built on Mistral’s Magistral. Building on a European base is the honest version.

The EU consortia. EuroLLM released a 22B model in December 2025 and is the one shipping. OpenEuroLLM published experimental 9B checkpoints in August and is still training a 9B base called Prelude, but it has no flagship. The Domyn-led EUROPA consortium was picked on 19 June to build a 400 billion parameter model in all 24 EU languages, with 2.5% of EuroHPC capacity for a year: a promise with a compute allocation attached, not a model. The Commission’s translation service took a fourth route: Mistral’s Mixtral 8x7B continued on institutional text, published in July, Apache 2.0 but downloadable only inside the EU.

Who Actually Downloads European LLMs?

We pulled last-30-day download counts from the Hugging Face API on 6 October 2026, using each model’s most pulled official repository.

How Reliable Are These Download Numbers?

• Downloads are not usage. API and self-hosted traffic do not show up, which is why Large 3 looks small.
• Builds people can run win. Mellum2’s Q8_0 build pulls 21 times its full-weight repository, Bielik’s AWQ almost seven times.
• Bulk pulls distort. NorwAI’s gated Norwegian model read 542,008 in September and 100,485 now, which is what an automated pull looks like once it stops.
• Hugging Face restates the counts. Bielik’s AWQ build read 218,127 in September and 63,980 now: treat every figure as a snapshot.

Europe’s most downloaded open LLM is still a 7B Mistral released in May 2024. Behind it the order changed: Apertus 1.5 went from 36,520 downloads to 697,138 after its 17 September release, putting a Swiss public model second, ahead of Poolside and JetBrains.

The split is not public against private. Spain shows both shapes in one programme: the Barcelona Supercomputing Centre’s small Salamandra pulls tens of thousands a month, while its €10.2M ALIA-40B pulls 2,545. Mistral’s 675B release pulls less than its 2024 7B does in a day, and Kolibri needs 78GB of memory even in FP8. What decides adoption is not who paid for the model, but whether it fits the hardware the reader already owns.

How Do European LLMs Compare to Chinese Models?

One test for both maps: in 2026, did the lab release open weights for a model of its own above 100 billion parameters? Two pass here: Mistral, with Small 4 and Medium 3.5, and Poolside, with Laguna M.1 and S 2.1. Applying the same test to our Chinese LLM map, fourteen Chinese labs do. Large 4’s 1 trillion parameter weights are due on 27 October; until then Europe’s largest released open-weight model is Large 3 at 675B, against Moonshot’s Kimi K3 at 2.8 trillion.

On the overall leaderboard the gap is wider. Mistral’s best entry, Medium 3.5, ranked 115th of 413 models on 6 October.

The compute gap explains most of it. Mistral raised $830M in March for a data centre south of Paris with 13,800 Nvidia GB300 chips. Meta expects $130bn to $145bn of capital expenditure this year. The EU’s answer, the AI Gigafactories call launched on 30 July with up to €10bn, mostly arrives in 2027 and later, the wrong side of the gap.

What to Watch for the Rest of 2026

• What licence Large 4 carries. The weights land on 27 October and Mistral has not said under what terms. Small 4 is Apache 2.0; Medium 3.5 has a revenue cap.
• Whether Cohere keeps Aleph Alpha’s models open. The agreement was signed on 16 September, closing still needs regulators, and Kolibri shipped Apache 2.0 two weeks later.
• What Nvidia does with European open models. It has agreed to buy Hugging Face, where every number above is counted, for $12.93bn, and reportedly licensed Poolside’s factory for $6bn.
• The EUROPA 400B model. Nothing yet, against an open-weight leader at 2.8 trillion.

Europe has more working models than the “no AI” line suggests, and fewer than the announcements imply. We test them one by one; the reviews linked here are where the counting gets done.

Author: Akos Szima

This article is for information only and is not financial or legal advice. First published 29 July 2026 and updated 25 August, 16 September and 6 October 2026. Download counts are last-30-day Hugging Face API figures from 6 October 2026, a snapshot that Hugging Face restates. Licence summaries are ours: read the licence file before you deploy.

See Also:

Chinese LLMs in 2026: The Complete Map

What Is the Best Open-Source LLM in 2026?

The Best Local LLM Nobody Can Run: Apertus 1.5 Measured

What Is Bielik? Poland’s Sovereign AI Model, Tested (2026)

Which VCs Are Actually Backing European AI in 2026?