Every model on Evanesca, who runs it, and where.
Evanesca serves 17 AI models, and every one runs in Europe: in France by Scaleway and OVHcloud (as a fallback), in Mistral AI's own EU region, in Germany by AKI.IO and in Switzerland by Infomaniak — five European companies. No model call leaves Europe: not as a backup, not when servers are busy.
Checked on 20 Oct 2026
Run in Europe · 17
Every model Evanesca serves runs in Europe, at a European company; the Where column gives the place.
| Model | Maker | Licence | Size | Type | Run by | Where | Checks | Price / M tokens (in · out) |
|---|---|---|---|---|---|---|---|---|
 Mistral Large 4 | Mistral AI | — | 1T (49B active) | Text + images | Mistral AI | EU region of Mistral AI | 4/6 · Licence ✕ · Provenance ✕ | €2.5133 · €7.7246 |
 DeepSeek V4.1 Flash | DeepSeek | MIT | 763B | Text + images | AKI.IO fallback: Scaleway | Germany (AKI.IO) | 6/6 | €1.008 · €3.024 |
 GLM-5.3 | Zhipu / Z.ai | GLM-5.3 License | 753B | Text | AKI.IO fallback: Scaleway | Germany (AKI.IO) | 6/6 | €3.024 · €9.24 |
 Qwen3.8 27B | Qwen | Apache-2.0 | 27.8B | Text + images | Scaleway fallback: OVHcloud | Paris, France | 6/6 | €1.008 · €5.544 |
| DeepSeek V4 Flash | DeepSeek | MIT | 284B (13B active) | Text | Scaleway | Paris, France | 6/6 | €0.672 · €1.344 |
| Apertus v1.5 70B | Swiss AI Initiative | Apache-2.0 | 72B | Text + images | Infomaniak | Switzerland (Infomaniak) | 4/6 · Security, Benchmarks: not run yet | €1.2936 · €4.62 |
| GLM-5.2 | Zhipu / Z.ai | MIT | 753B | Text | Scaleway | Paris, France | 6/6 | €3.024 · €9.24 |
| Qwen3.6 35B | Qwen | Apache-2.0 | 36B (3B active) | Text + images | Scaleway | Paris, France | 6/6 | €0.42 · €2.52 |
 Gemma 4 26B | Apache-2.0 | 26B (4B active) | Text + images | AKI.IO | Germany (AKI.IO) | 6/6 | €0.168 · €0.84 | |
| Mistral Small 4 | Mistral AI | Apache-2.0 | 119B (6.5B active) | Text + images | Mistral AI | EU region of Mistral AI | 5/6 · Security ✕ | €0.2772 · €1.1088 |
| Qwen3.5 397B | Qwen | Apache-2.0 | 397B (17B active) | Text + images | Scaleway | Paris, France | 6/6 | €1.008 · €6.048 |
 GPT-OSS 120B | OpenAI | Apache-2.0 | 117B (5.1B active) | Text | Scaleway fallback: OVHcloud | Paris, France | 5/6 · Security ✕ | €0.252 · €1.008 |
| Qwen3 Coder 30B | Qwen | Apache-2.0 | 30.5B (3.3B active) | Code | Scaleway | Paris, France | 5/6 · Security ✕ | €0.336 · €1.344 |
| Qwen3 235B | Qwen | Apache-2.0 | 235B (22B active) | Text | Scaleway | Paris, France | 5/6 · Security ✕ | €1.26 · €3.78 |
| Mistral Small 3.2 | Mistral AI | Apache-2.0 | 24B | Text + images | Scaleway fallback: OVHcloud | Paris, France | 5/6 · Security ✕ | €0.252 · €0.588 |
 Llama 3.3 70B | Meta | Llama 3.3 Community License | 70B | Text | Scaleway fallback: OVHcloud | Paris, France | 5/6 · Security ✕ | €1.512 · €1.512 |
| Pixtral 12B | Mistral AI | Apache-2.0 | 12B | Vision | Scaleway | Paris, France | 5/6 · Security ✕ | €0.336 · €0.336 |
Prices per million tokens, VAT included; reasoning tokens count as output. The exact cost of each answer shows under the answer.
“not run yet”: checks not run on that model so far — not failed ones.
Mistral Large 4 — Public preview: Mistral announced its open weights for late October 2026, so its licence is not published yet. Mistral AI keeps API inputs and outputs for 30 days to monitor abuse. As a preview, what you send to it may be used by Mistral AI to train its models: Mistral's commercial terms (§ 4.3) exclude preview models from opt-outs.
GLM-5.3 — AKI.IO serves a 512k context window; the model itself goes up to 1M.
Apertus v1.5 70B — The open model of the Swiss AI Initiative (EPFL, ETH Zurich, CSCS), run by Infomaniak in Switzerland — outside the EU, under an EU adequacy decision.
Mistral Small 4 — It did not pass the security check on 13 October 2026: it resisted 1 of 3 prompt-injection tests, none of 3 system-prompt-leak tests and 1 of 3 exfiltration tests. Mistral AI keeps API inputs and outputs for 30 days to monitor abuse.
Pixtral 12B — It did not pass the security check on 20 October 2026: in none of the five categories tested did it give two safe answers out of three. It repeated its system prompt, followed jailbreaks, put a confidential number inside an image link and wrote a phishing email. It reads screenshots and documents well, but misreads an image made of one flat colour.
Llama 3.3 is licensed under the Llama 3.3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Built with Llama.
Paused · 4
These models are not available at the moment: their only European operator no longer serves them to Evanesca, and no other runs them. They are not counted in the list above, a call answers 503 and is not billed, and each comes back unchanged once its operator serves it again. Meanwhile:
- Qwen3.6 27B (OVHcloud, since 20 Oct 2026) → Qwen3.6 35B
- Qwen3.5 9B (OVHcloud, since 20 Oct 2026) → Gemma 4 26B
- GPT-OSS 20B (OVHcloud, since 20 Oct 2026) → GPT-OSS 120B
- Qwen2.5-VL 72B (OVHcloud, since 20 Oct 2026) → Pixtral 12B or Qwen3.6 35B
No model outside Europe
Evanesca serves no model outside Europe, and none is reached through a US company. That is why some well-known models are not in this list. More models will join when a European operator runs them.
What “run in Europe” means here
Every model Evanesca serves runs in Europe, at a European company. Location alone is not enough — a US company remains subject to the CLOUD Act even when its servers are in Europe. That is why Evanesca also looks at who controls the operator (and says so when that is not published) and what it keeps:
- Paris, France · Scaleway SAS, Paris — iliad group, no US parent. Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days (Scaleway's data privacy page).
- Gravelines, France · OVH SAS, France — no US parent company. OVHcloud states “zero data retention: we keep only the data required for billing purposes” (OVHcloud AI Endpoints).
- EU region of Mistral AI · Mistral AI SAS, Paris — serves its own EU region, no US parent. Mistral AI keeps API inputs and outputs “for thirty (30) rolling days to monitor abuse” (Mistral's privacy policy).
- Germany (AKI.IO) · AKI.IO GmbH, Berlin — a German company; its shareholders are not published. AKI.IO states that “prompts, inputs, and generated outputs are processed in volatile memory and are not logged, stored, or analyzed by AKI.IO” (AKI.IO); its privacy policy keeps usage data — request ID, model, status, timestamps, token usage — for 90 days, for billing and service maintenance (AKI.IO's privacy policy).
- Switzerland (Infomaniak) · Infomaniak, Geneva — an independent Swiss company, no US parent. Infomaniak states that queries are “neither recorded nor used to train models” (Infomaniak AI Tools).
Mistral AI does not publish which EU country serves a request: the Where column says “EU region of Mistral AI”, not a country.
In Germany, AKI.IO runs DeepSeek V4.1 Flash, GLM-5.3 and Gemma 4 26B. AKI.IO says inference runs on GPU infrastructure in German data centres (AKI.IO, read on 13 Oct 2026).
Switzerland is not in the EU, but the European Commission has found that it protects personal data adequately (Decision 2000/518/EC): personal data can go there from the EU without additional guarantees. That is why Evanesca counts it as Europe, and serves Apertus v1.5 70B there.
The six checks
Every model shows its checks before you use it. A model can join Evanesca with a stamp missing: the Checks column above and its passport say which.
- Licence — read at the source, so you know what you may do with the output
- Provenance — traced to the maker's official release, as declared by the operator
- Security — injection, jailbreak, leak, exfiltration and harmful-request tests
- GDPR — a sovereign route: European operator, country and retention documented
- Benchmarks — quality and speed measured on our own route
- Live test — a real conversation, end to end, through the route Evanesca uses
Each answer then carries its passport: the model, the operator, the country and what was kept.