Models and Europe

Every model Evanesca serves 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. No model call leaves Europe, and every response says who served it and where.

Catalogue checked on 20 Oct 2026

SoonOpening soon: what follows describes the service as it will work.

Where each model runs

No model call leaves Europe: not as a backup, not when servers are busy. If an operator fails, you get its error, not an answer from somewhere else. An ID Evanesca does not serve answers 404 model_not_found, and the request goes nowhere else.

The model field takes the IDs below. GET /models returns the same list as JSON, with no key needed; it holds chat, embedding and image models, and each entry's kind says which (add ?kind=chat, ?kind=embedding or ?kind=image to get one family). GET /models/{id} returns a single entry (the reference).

Chat models
Model	ID	Type	Context	Tool calling	Reasoning	Runs in
Mistral Large 4	`mistral-large-4`	Text + images	512k	Yes	—	EU region of Mistral AI
DeepSeek V4.1 Flash	`deepseek-v4.1-flash`	Text + images	1M	Yes	Yes	Germany (AKI.IO)
GLM-5.3	`glm-5.3`	Text	512k	Yes	Yes	Germany (AKI.IO)
Qwen3.8 27B	`qwen3.8-27b`	Text + images	256k	Yes	Yes	Paris, France
DeepSeek V4 Flash	`deepseek-v4-flash`	Text	256k	Yes	Yes	Paris, France
Apertus v1.5 70B	`apertus-v1.5-70b`	Text + images	100k	Yes	—	Switzerland (Infomaniak)
GLM-5.2	`glm-5.2`	Text	256k	Yes	Yes	Paris, France
Qwen3.6 35B	`qwen3.6-35b`	Text + images	256k	Yes	Yes	Paris, France
Gemma 4 26B	`gemma-4-26b`	Text + images	256k	Yes	—	Germany (AKI.IO)
Mistral Small 4	`mistral-small-4-119b`	Text + images	256k	Yes	—	EU region of Mistral AI
Qwen3.5 397B	`qwen3.5-397b`	Text + images	250k	Yes	Yes	Paris, France
GPT-OSS 120B	`gpt-oss-120b`	Text	—	Yes	Yes	Paris, France
Qwen3 Coder 30B	`qwen3-coder-30b`	Code	256k	Yes	—	Paris, France
Qwen3 235B	`qwen3-235b`	Text	250k	Yes	—	Paris, France
Mistral Small 3.2	`mistral-small-3.2`	Text + images	—	Yes	—	Paris, France
Llama 3.3 70B	`llama-3.3-70b`	Text	—	Yes	—	Paris, France
Pixtral 12B	`pixtral-12b`	Vision	128k	Yes	—	Paris, France

Context: as the operator announces it; “—” means not checked yet, never “none”. Tool calling: measured with a real call at the operator, not read on a model card. Reasoning: models flagged as thinking before they answer. Prices per million tokens are on the models page and under billing.

Every model in the table returns tool calls.

The embedding models run in France (Embeddings); the image models in Switzerland, at Infomaniak (Images).

The operators

An operator runs the model: it is the company that receives the prompt. There are 5, all European, and the API names the one that served each response.

The API's operators
Operator	Company	Where	Models
Mistral AI	Mistral AI SAS, Paris — serves its own EU region, no US parent	EU region of Mistral AI	`mistral-large-4`, `mistral-small-4-119b`
AKI.IO	AKI.IO GmbH, Berlin — a German company; its shareholders are not published	Germany	`deepseek-v4.1-flash`, `glm-5.3`, `gemma-4-26b`
Scaleway	Scaleway SAS, Paris — iliad group, no US parent	Paris, France	`qwen3.8-27b`, `deepseek-v4-flash`, `glm-5.2`, `qwen3.6-35b`, `qwen3.5-397b`, `gpt-oss-120b`, `qwen3-coder-30b`, `qwen3-235b`, `mistral-small-3.2`, `llama-3.3-70b`, `pixtral-12b`, `qwen3-embedding-8b`, `bge-multilingual-gemma2`
OVHcloud	OVH SAS, France — no US parent company
Infomaniak	Infomaniak, Geneva — an independent Swiss company, no US parent	Switzerland	`apertus-v1.5-70b`, `flux-schnell`, `photomaker-v2`

AKI.IO's place is what AKI.IO states: “Inference requests are currently processed on GPU infrastructure in German data centers” (AKI.IO, read on 13 Oct 2026).

What the operators keep

Evanesca stores no prompt and no answer: they stay in memory while the answer is written. Mistral AI keeps API inputs and outputs “for thirty (30) rolling days to monitor abuse” (Mistral's privacy policy). 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). Scaleway applies zero data retention by default; when requests cause errors or look abusive, it may keep and read their content for up to 14 days (Scaleway). OVHcloud states “Zero data retention: We keep only the data required for billing purposes” (OVHcloud). Infomaniak, which also runs the image models, states that queries are “neither recorded nor used to train models” (Infomaniak).

Every response says where it was served

Every chat, embeddings or image response that reached an operator, streamed or not, carries two headers:

Response header	Value
`x-evanesca-operator`	The operator that served the call: mistral, aki, scaleway, ovhcloud or infomaniak.
`x-evanesca-location`	Where it ran: paris (Paris, France), gravelines (Gravelines, France), mistral-eu-region (EU region of Mistral AI), germany (Germany) or switzerland (Switzerland).

curl -i

curl -i https://evanesca.app/api/v1/chat/completions \
  -H "Authorization: Bearer $EVANESCA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gemma-4-26b", "messages": [{"role": "user", "content": "Hello"}]}'

Response headers (excerpt)

HTTP/2 200
content-type: application/json
x-evanesca-operator: aki
x-evanesca-location: germany

Each model's compliance record

Besides OpenAI's fields (id, object, created, owned_by), each GET /models entry carries what you need to choose — context_length and tool_calling when they have been checked, reasoning, and pricing in euros per million tokens, VAT included — and the model's full compliance record:

Field	What it says
`operator, location`	Who runs the model, and where — the values of the two response headers: mistral, aki, scaleway, ovhcloud or infomaniak; paris, gravelines, mistral-eu-region, germany or switzerland.
`licence, open_weights`	The licence as published — null while none is — and whether the weights are public.
`checks`	The six checks as booleans — licence, provenance, security, gdpr, benchmarks, live_test — and, in run_on, the date of each passed check where it is known. false means missing: not run yet, or run without meeting the pass mark (how each check is run).
`retention`	What the operator says it keeps (rule), and the page where it says so (source). Evanesca itself stores no prompt.
`training`	What the operator says about training on your data (statement), and its source (source).

Embedding and image entries carry operator, location, retention and training too. Here is the full entry of Gemma 4 26B, built from the catalogue:

GET /models/gemma-4-26b · 200 · application/json

{
  "id": "gemma-4-26b",
  "object": "model",
  "kind": "chat",
  "created": 1792454400,
  "owned_by": "Google",
  "operator": "aki",
  "location": "germany",
  "context_length": 256000,
  "tool_calling": true,
  "reasoning": false,
  "pricing": {
    "currency": "EUR",
    "vat_included": true,
    "input_per_million": 0.168,
    "output_per_million": 0.84
  },
  "licence": "Apache-2.0",
  "open_weights": true,
  "checks": {
    "licence": true,
    "provenance": true,
    "security": true,
    "gdpr": true,
    "benchmarks": true,
    "live_test": true,
    "run_on": {
      "security": "2026-10-13",
      "benchmarks": "2026-10-13",
      "live_test": "2026-10-13"
    }
  },
  "retention": {
    "rule": "Inference content “is processed only in volatile memory […] and is not logged, stored, or analyzed by AKI.IO”; usage data (request ID, model endpoint, status, timestamps, token usage) “is retained for 90 days”.",
    "source": "https://aki.io/privacy-policy"
  },
  "training": {
    "statement": "“Inference requests are not forwarded to model developers, and customer prompts and outputs are not used to train AI models.”",
    "source": "https://aki.io/"
  },
  "status": "live"
}

Every field, and the call in four languages: retrieve a model. What each check means: transparency.