LiteLLM on AI models run in Europe

Setup

import os
import litellm

response = litellm.completion(
    model="openai/gemma-4-26b",
    api_base="https://evanesca.app/api/v1",
    api_key=os.environ["EVANESCA_API_KEY"],
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

model_list:
  - model_name: gemma-4-26b
    litellm_params:
      model: openai/gemma-4-26b
      api_base: https://evanesca.app/api/v1
      api_key: os.environ/EVANESCA_API_KEY
    model_info:
      # Evanesca's price, € per token, margin and VAT included
      input_cost_per_token: 0.000000168
      output_cost_per_token: 0.00000084
  - model_name: bge-multilingual-gemma2
    litellm_params:
      model: openai/bge-multilingual-gemma2
      api_base: https://evanesca.app/api/v1
      api_key: os.environ/EVANESCA_API_KEY
    model_info:
      mode: embedding
      input_cost_per_token: 0.000000168

litellm_settings:
  drop_params: true

export EVANESCA_API_KEY="<your Evanesca key>"
export LITELLM_MASTER_KEY="sk-$(openssl rand -hex 24)"
litellm --config config.yaml --port 4000

stream = litellm.completion(
    model="openai/gemma-4-26b",
    api_base="https://evanesca.app/api/v1",
    api_key=os.environ["EVANESCA_API_KEY"],
    messages=[{"role": "user", "content": "Bonjour"}],
    stream=True,
    stream_options={"include_usage": True},
)

Which model to pick

Evanesca's models from which a real tool call came back at their operator. This guide uses DeepSeek V4.1 Flash and Gemma 4 26B, picked from the catalogue for their properties. Prices per million tokens, input / output, margin and VAT included.

Reasoning tokens are billed as output. Where each model runs and what its operator keeps: Models and Europe.

Good to know

• Clients that call /v1/responses (OpenAI's newer format) fail with a 404 behind the openai/ prefix, since Evanesca serves Chat Completions. LiteLLM documents a custom_openai/ prefix that bridges that route into Chat Completions; keep openai/ for embedding models.
• LiteLLM does not convert currencies: it multiplies tokens by the cost you enter. With Evanesca's prices, the spend it tracks is in euros — mind it if the same budget also covers models priced in dollars.
• If the same proxy also routes to providers outside Europe, LiteLLM's fallbacks can move a request from one to the other. Keep the fallback chains of regulated workloads inside Evanesca's models.

Where your prompts go

LiteLLM runs where you deploy it: the SDK in your process, the proxy on your server. It sees every prompt before Evanesca does, and its logs and database are yours to place.

On Evanesca's side, the model runs at a European operator, named with its place in the x-evanesca-operator and x-evanesca-location headers of every response, and Evanesca keeps no copy of prompts or answers — only the billing data of each call. What each operator keeps is set out in Models and Europe; the host of Evanesca's servers is named in the legal notice.

Troubleshooting

Sources

Official documentation, read on 12 Oct 2026:

• LiteLLM — OpenAI-compatible endpoints
• LiteLLM — Proxy config.yaml
• LiteLLM — Custom pricing
• LiteLLM — Usage in streams
• LiteLLM — drop_params
• LiteLLM — /v1/responses and /v1/messages passthrough

Frequently asked questions