Introduction
I have written about Europe’s AI sovereignty problem four times now (Green 2026a, 2026b), and every time I land in the same uncomfortable place. The two pillars Europe leans on (the closed US frontier and the open Chinese weights frontier) are both rotting, and the continent I live in still has no pillar of its own.
The snapshot date above the table is the version. It changes when a frontier model lands, the AA index changes version, or a set of weights moves behind an endpoint.
The two pillars
Europe runs its frontier AI on two pillars we do not control. Pillar one is the closed US frontier (Anthropic, OpenAI, Meta, xAI, Google) accessed through APIs whose terms can change overnight based on a government we have nothing to do with. Pillar two is the open Chinese weights frontier (Kimi, Xiaomi’s MiMo, DeepSeek, the Qwen open line, and the GLM line) which ships real weights today but ships them from a jurisdiction that published draft export controls on model weights on 2026-10-16 (The Next Web 2026; The Decoder 2026). Both pillars can be pulled. We saw pillar two wobble in slow motion when Qwen3.8-Max launched API-only with weights “promised within days”, watched it partly reset on 2026-08-13 when the 2.4T-A95B weights landed on HuggingFace, watched it wobble again when Z.ai shipped GLM-5.3 API-first and delayed the weights over the model’s own cyber capability, and then watched the promise kept in full: the Flash tier landed open on 2026-08-26, and the full 5.3 weights landed on 2026-08-28, two weeks to the day after the API launch, on the schedule Z.ai named up front (Z AI 2026a, 2026b). We saw pillar one’s political risk go from theoretical to concrete when Anthropic tried corporate self-regulation for military use and got punished so publicly that no rational actor will try it again (Anthropic 2026). Europe has its own version of the rented pillar. Spain’s Multiverse Computing serves a compressed, proprietary copy of GLM-5.2 from its own API.
What changed
Europe kept a promise, and Artificial Analysis replaced its yardstick.
Mistral published the Large 4 weights on Hugging Face on 2026-10-29, inside the end-of-October window it named at the preview, under a modified MIT license (Mistral AI 2026). Europe now has a 1-trillion-parameter model that it trained in its own datacenters and can run on its own hardware. StepFun kept its date too, with the Step 5 Preview weights on Hugging Face and ModelScope on 2026-10-15 (Pandaily 2026).
The index moved to v5 on 2026-10-27 (heise online 2026). It is a new scale, and it moved the open models down as a group. Mistral Large 4 Preview graded 38.4 on v4.3.2 and grades 25.3 on v5. GLM-5.3 went from 44.8 to 31.7, MiMo-V2.6-Pro from 46.3 to 29.9, Step 5 Preview from 43.7 to 28.5, and Qwen3.8-27B from 33.7 to 20.1. Over the same rebuild Claude Opus 5.5 went from 57.6 to 59.1 and GPT-6 Astra from 52.7 to 56.9. The same open models lost 13 to 16 points each, and the two closed leaders gained 1.5 and 4.2. Europe’s gap to the frontier went from 19.2 points to 38.5, and most of that move is the yardstick.
Two closed launches set the new top. Claude Fable 5.2, released 2026-11-09, grades 64.6, 5.5 points above Opus 5.5 (9to5Mac 2026). Anthropic shipped it alongside Mythos 5.2, a sibling with restricted availability. Gemini 4 Carbon, released 2026-10-28, grades 60.0 and went generally available on 2026-11-10, with an EU region on Google Cloud and its cyber variant behind an allowlist (Google Cloud 2026). Carbon replaces Argon as Google’s row, so the board’s only “Restricted” verdict is gone. GPT-6 Astra holds third at 56.9. Meta’s new closed flagship, Muse Wonder Full, grades 50.1 and falls out of the window, which now holds three labs.
Moonshot shipped Kimi K3.1 Pro and published its weights on Hugging Face in early November, under a modified MIT license with extra conditions for very large operators (AILog 2026). It has 2.8 trillion parameters, grades 46.5, sits 18.1 points under the top, and leads the next open model (DeepSeek V4.1 Flash 1017, 32.6) by 13.9 points. Meta is back in the open line too. Muse Spark Open, under a license that allows commercial use up to $20 million in revenue, grades 24.9, 1.2 points under Mistral (TechRepublic 2026).
Beijing turned its consultation into a draft rule. Announcement No. 48, issued on 2026-10-16 by the Ministry of Commerce and the Ministry of Science and Technology, proposes restricting the export of weights for any model trained with 10^26 FLOPs, with 2 trillion parameters, or ranked among the world’s ten leading models (The Decoder 2026). Weights already open-sourced are exempt, and serving a model through an API does not count as an export. Overseas on-premise deployment does. Comments close on 2026-11-15, and an interim registration requirement has applied since 2026-10-16. Kimi K3.1 Pro is above the parameter threshold, and its weights went out while the rule is still a draft.
DeepSeek V4.1 Pro has still not shipped (DeepSeek 2026), and the Qwen3.8-Max hybrid weights are fourteen weeks past “within days” (Alibaba 2026b). Multiverse’s Quasar 438B and Swiss AI’s Apertus 70B have no v5 grade yet.
The scoreboard
Snapshot date: 2026-11-11. Scores from the Artificial Analysis Intelligence Index v5, pulled from their Data API (free tier) (Artificial Analysis 2026b, 2026a). Each row is the top reasoning variant of that lab’s frontier model, selected automatically: a lab enters when its best model is within ten points of the board’s top score, European labs enter regardless of score, and the curated open-line rows carry the open-weights argument. The Δ top column is each row’s distance from the board’s best score. Treat every number as a point-in-time reading. Index v5 is a new scale, so no score on this page compares with an earlier snapshot, and rows marked TBD have no v5 grade yet.
| Rank | Lab | Region | Latest frontier model | Open? | AA index | Δ top | EU access | Verdict |
|---|---|---|---|---|---|---|---|---|
| 1 | Anthropic | US | Claude Fable 5.2 | Proprietary | 64.6 | 0.0 | API | Revocable (ToS and political risk) |
| 2 | US | Gemini 4 Carbon | Proprietary | 60.0 | −4.6 | API (GA 2026-11-10, EU region) | Revocable (cyber variant behind an allowlist) | |
| 3 | OpenAI | US | GPT-6 Astra | Proprietary | 56.9 | −7.7 | API | Revocable |
| 4 | Moonshot | China | Kimi K3.1 Pro | Open weights (Kimi K3.1 Pro license, modified MIT, shipped 2026-11) | 46.5 | −18.1 | Self-host | Revocable (Beijing draft export controls, Announcement No. 48) |
| 5 | Tencent | China | Tencent Hy4 | API-only; Hy4 Preview weights open (Apache 2.0), full Hy4 weights promised, no date | 34.6 | −30.0 | API | Revocable (Beijing draft export controls, Announcement No. 48) |
| 6 | Alibaba (Max) | China | Qwen3.8 Max | API-only (0902 refresh 2026-09-02); 2.4T-A95B sibling open; hybrid weights promised 2026-08-03, not shipped | 33.7 | −30.9 | Self-host (2.4T) · API (Max) | Revocable (2.4T open, Max API-only, hybrid weights promised 2026-08-03, not shipped) |
| 7 | DeepSeek | China | DeepSeek V4.1 Flash 1017 | Open weights | 32.6 | −32.0 | Self-host | Revocable (Beijing draft export controls, Announcement No. 48) |
| 8 | Z.ai | China | GLM-5.3 | Open weights (custom permissive license, shipped 2026-08-28) | 31.7 | −32.9 | Self-host | Revocable (Beijing draft export controls, Announcement No. 48) |
| 9 | Xiaomi | China | MiMo-V2.6-Pro | Open weights (MIT, shipped 2026-09-21) | 29.9 | −34.7 | Self-host | Revocable (Beijing draft export controls, Announcement No. 48) |
| 10 | StepFun | China | Step 5 Preview | Open weights (StepFun Model License 1.0, shipped 2026-10-15) | 28.5 | −36.1 | Self-host | Revocable (Beijing draft export controls, Announcement No. 48) |
| 11 | Mistral | EU | Mistral Large 4 | Open weights (Mistral Ecosystem Open-Weight license, modified MIT, shipped 2026-10-29) | 26.1 | −38.5 | Self-host | EU pillar, trained in Europe, self-hostable |
| 12 | Meta (open line) | US | Muse Spark Open | Open weights (Muse Model license, commercial use up to $20M revenue, shipped 2026-10-15) | 24.9 | −39.7 | Self-host | Revocable (Meta license, revenue-capped) |
| 13 | Alibaba (27B open) | China | Qwen3.8-27B | Open weights (Apache 2.0) | 20.1 | −44.5 | Self-host | Revocable (Beijing draft export controls, Announcement No. 48) |
| 14 | OpenAI (open line) | US | gpt-oss-120b | Open weights | 3.8 | −60.8 | Self-host | Stable, but far from frontier |
| n/a | Multiverse Computing | EU | Quasar 438B | Proprietary (AA: based on GLM-5.2) | TBD | TBD | API | EU pillar, under construction, far from frontier (AA: GLM-5.2 derivative); not graded on the current index |
| n/a | Swiss AI Initiative | Europe (non-EU) | Apertus 70B Instruct | Open weights | TBD | TBD | Self-host | European, outside EU jurisdiction; not graded on the current index |
| n/a | EUROPA / Domyn | EU | In training | TBD | TBD | TBD | EU native | Under construction, >400B params, 24 EU languages |
The Qwen family spans three openness states: a closed Max endpoint (33.7 for the 0902 refresh), an open 2.4T sibling, and an Apache 2.0 consumer 27B (20.1). The 2.4T and the Flash-Next architecture preview have no v5 grade yet, and the Max hybrid is fourteen weeks past its launch-day promise (Alibaba 2026b, 2026a). Tencent runs the same split one tier down. The Hy4 Preview weights shipped under Apache 2.0 on 2026-08-28, and the full Hy4, released 2026-11-03, is API-only so far (South China Morning Post 2026). The Multiverse cell is proprietary, based on GLM-5.2, and served from Spain (Artificial Analysis 2026d). Meta’s Muse Spark Open, at 24.9, is now the best American open-weights row, and OpenAI’s gpt-oss-120b grades 3.8.
How to read it
The AA index moves week to week as it re-aggregates, and it changes scale when the yardstick gets replaced. Index v5, released 2026-10-27, aggregates nine evaluations in four equally weighted categories (knowledge work, coding, science, general). It added FrontierCode 2.0 and Terminal-Bench-Science, replaced Humanity’s Last Exam with HLE-Diamond, updated AA-Omniscience and CritPt, and dropped GDPval-AA, SciCode and AA-LCR (heise online 2026). Roughly 60 percent of the composite is now graded on private or held-out test sets. A model can gain or lose a point or two without a new release, just because the index re-sampled, and it can lose fifteen when the yardstick changes. So a point or two between the rows at the top of the board is noise. The signal is which labs are within shouting distance of the frontier top, and which access regime they ship under. The frontier window now closes after row 3, GPT-6 Astra at 56.9. Muse Wonder Full (50.1), Kimi K3.1 Pro (46.5), Grok 4.7 (35.5) and Tencent Hy4 (34.6) all sit below it.
“Open?” is the column that matters most for Europe, and it is the one most likely to lie to you. This round three labs named dates for weights and kept them. StepFun shipped Step 5 on 2026-10-15, Mistral shipped Large 4 on 2026-10-29, and Moonshot published Kimi K3.1 Pro in early November (Pandaily 2026; Mistral AI 2026; AILog 2026). The Qwen cell is the standing exception. The Max hybrid was promised “within days” on 2026-08-03, and fourteen weeks later the model card is still not on HuggingFace, while everything below the Max line is open (Alibaba 2026b). The 27B cell has no promise and no asterisk: Apache 2.0, on HuggingFace, downloadable today.
“EU access” comes in two kinds again. API access means you can call the model today and you cannot run it yourself tomorrow if the provider decides you should not. Self-host means you have the weights on hardware you control, and the only thing that can take that away is a license change you can see coming or a jurisdiction deciding to restrict exports. Argon’s Fairwind-only access left the board when Carbon became Google’s best model, and Carbon reaches Europe through Google Cloud’s EU region (Google Cloud 2026). Beijing’s draft draws the line between the two kinds. Overseas on-premise deployment of a covered model counts as an export and API serving does not (The Decoder 2026). For a covered model whose weights are not already public, the draft leaves Europe the API and takes away the self-host.
Where Europe sits
Mistral is Europe’s best-graded lab, with Large 4 ranked #72 among all graded variants (Artificial Analysis 2026c), and the first European model on this board with a trillion parameters and downloadable weights. The ASML-led 1.7 billion euro Series C put real compute behind the company (CNBC 2025). On v5 Large 4 sits 38.5 points under the top and 20.4 points under Kimi K3.1 Pro, the best open model. The frontier-tier question (does Mistral ship a model that sits in the top three of the AA index, ever) is still open. Europe can now run its best model on its own hardware, and that model trails five Chinese open-weights rows.
Multiverse Computing has no v5 grade. Quasar 438B (26.7 on v4.3.2) is proprietary, served from Multiverse’s own CompactifAI API, and, by Multiverse’s own technical note, “a compressed model built from GLM-5.2”, the Chinese open-weights model, pruned with their quantum-inspired method (Multiverse Computing 2026; Artificial Analysis 2026d). At launch Multiverse billed it as Europe’s leading AI model, and five weeks later Mistral’s own model passed it. On v4.3.2 the open GLM-5.2 original led the compressed copy 33.7 to 26.7, and the original is free to self-host while the copy is API-only. Their previous graded model, HyperNova 60B, is the same pattern: AA labels it “based on gpt-oss-120b”, the American open-weights artifact. Both of Multiverse’s graded rows are compressions of other labs’ open models. That is a real business, and it is a long way from a frontier.
The EUROPA consortium led by Domyn is the other one to watch, and it is unchanged. The EU Frontier AI Grand Challenge is funding a >400 billion parameter model trained on all 24 EU languages on a 6,000-chip Nvidia Blackwell cluster (European Commission 2026). It is still the only announced European project that names the frontier tier as its goal. It is also not a frontier model yet, and the gap between “we are training a 400B model” and “we have a model in the top three of the AA index” is exactly the gap Europe has been failing to close for two years.
Apertus 70B (Swiss AI Initiative) is real, downloadable and European, and it has no v5 grade either. On v4.3.2 it graded 5.1, 52.5 points off the frontier.
There is still no European row in the top half of the table. Mistral Large 4 is row 11 of 14 scored rows.
What would change the table
These would move the table beyond normal score drift:
- Beijing finalises Announcement No. 48 after comments close on 2026-11-15. Kimi K3.1 Pro already clears the 2-trillion-parameter test, and the final text decides whether the next Chinese model of that size ships open. If it does not, the Chinese verdicts flip from “Revocable” to “Restricted” and pillar two gets measurably weaker for Europe (The Decoder 2026).
- DeepSeek V4.1 Pro ships. DeepSeek’s best graded model is V4.1 Flash 1017, open weights, at 32.6. V4.1 Pro has not shipped, and V4 Pro stays on the API for now (DeepSeek 2026). If V4.1 Pro lands near Kimi, DeepSeek is back.
- Qwen3.8-Max hybrid weights actually ship. Fourteen weeks and one dated API refresh later, still no model card. Z.ai, StepFun, Mistral and Moonshot have all kept dated weights promises; Alibaba’s is the standing counterexample. I will believe the hybrid when the model card lands on HuggingFace.
- Qwen4 ships. Alibaba previewed it as in training at Apsara on 2026-09-22, with no dates (Alizila 2026). If Qwen4 lands as a Max-class open drop, the open top of the board is in play. If it lands API-only with open siblings below, the Qwen pattern repeats at the next generation.
- Tencent publishes the full Hy4 weights. The row flips from API-only to self-host, and Tencent joins the Chinese open frontier with its best model.
- A US provider changes terms in a way that breaks EU access (region lock, use-case restriction, government-mandated cutoff). The corresponding US row gets a date stamped on its verdict.
- EUROPA publishes a model, or Mistral lands one in the AA index top ten. A row leaves or joins the board.
- Meta or OpenAI ships an open-weights model near Kimi K3.1 Pro. Muse Spark Open put Meta back in the open line at 24.9. A US open model in the forties gives the two-pillar story a third leg I did not expect.
- A non-US lab re-enters the frontier window. Right now the window holds three American closed labs and nothing else. Kimi K3.1 Pro sits 8.1 points under the line.
Conclusion
Every time a frontier model lands, the good news is about someone else’s model and the bad news is about someone else’s decision. Europe’s role is to read the announcement and adjust its compliance documentation.
This round Europe kept a promise. Mistral named the end of October for the Large 4 weights and shipped them on 2026-10-29, so Europe’s best model is one it trained and can run on its own hardware. Thirteen of the fourteen scored rows are still non-European, and the new index puts the closed top 38.5 points above Mistral. The open frontier above it is Chinese, led by Kimi K3.1 Pro at 46.5, and Beijing now has a draft rule that decides whether the next model that size ships open. The closed US frontier is three labs, all revocable. Of the three pillars, the European one is the only one Europe controls, and it now has trillion-parameter weights in it.



