Managed inference · Impossible Cloud

Managed inference that fits how you work.
Three deployment models. One API.

For your apps and agents

Per token

For developers and the agents they ship. Pay only for the tokens you use.

Pay per token, no commitment

Built for coding agents and automation

Great for RAG pipelines and new products

Flexibility

Most flexible

Predictability

Performance

From €0.05 per 1M tokens

Start building

See the full price list

For your team

Per seat

For companies rolling out work AI. Your people chat with open models and your documents.

One flat price per person, per month

Buy direct or through your IT reseller

Popular in work AI apps like Plotdesk

Flexibility

Predictability

Most predictable

Performance

From €19 per seat/month2

Talk to sales

For production workloads

Dedicated

For customer-facing apps and high-volume pipelines with steady traffic.

GPUs reserved for you, around the clock

Predictable latency and throughput

Costs less per token at steady volume

Flexibility

Predictability

Performance

Best performance

Commitment pricing, agreed per contract

Talk to sales

Training or fine-tuning and need the whole machine?

See bare metal GPU

One API for all three.
Start per token and move to dedicated
without changing your code.

Any OpenAI-compatible tool works with every option.

Privacy

Your data stays yours

Private AI inference in European data centers.
Your prompts and responses are never stored or used for training.

Not cached

Your prompts and responses aren’t cached or logged. They exist for the request, then they’re gone.

Not retained

We don’t keep your data after the request completes. Nothing sits on our systems waiting to leak.

Not for training

Your data never trains our models, or anyone else’s. Ever.

Langdock

Open WebUI

n8n

Cline

Your own app

Europe Open models Runs in Europe Mistral, Qwen, DeepSeek, gpt-oss and more

Hosted in EuropeEuropean data centers, European law

No prompts storedProcessed, then gone

Not sure where to start? Talk to an expert

Tell us your use case and your budget, and our engineers help you pick the right model, size it, and get it running. No sales runaround, just people who know the stack.

Schedule a call

Marcel Jost

Technical Solutions Lead

Works with the tools you already use

Point any OpenAI-compatible app at Impossible Cloud. Only the base URL and the key change,
and your data stays in Europe.

Coding agents

Claude Code

Codex

Cline

Open Code

Kilo Code

Continue

Mistral Vibe

See docs ›

Automation

n8n

Make

LangChain

Dify

Activepieces

See docs ›

RAG and search

LlamaIndex

RAGflow

Haystack

Onyx

PipesHub

See docs ›

Work AI

Open WebUI

OpenClaw

Langdock

Plotdesk

See docs ›

Get started in your code

Any OpenAI client works. Copy, paste, run.

See docs

curl https://eu-west-3.inference.impossibleapi.net/v1/chat/completions \
  -H "Authorization: Bearer $IC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-27b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://eu-west-3.inference.impossibleapi.net/v1",
    api_key=os.environ["IC_API_KEY"],
)

reply = client.chat.completions.create(
    model="qwen3.8-27b",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://eu-west-3.inference.impossibleapi.net/v1",
  apiKey: process.env.IC_API_KEY,
});

const reply = await client.chat.completions.create({
  model: "qwen3.8-27b",
  messages: [{ role: "user", content: "Hello!" }],
});
console.log(reply.choices[0].message.content);

Full price list

Prices per 1M tokens, sorted cheapest first. All models hosted in Europe.

TEXT

CONTEXT

INPUT

OUTPUT

Nemotron 3 Nano 30B

262K

€0.05

€0.22

Nemotron 3.5 Lightning

€0.05

€0.22

Gemma 3 27B

110K

€0.09

€0.27

Qwen3 30B Instruct

262K

€0.09

€0.27

Qwen3 235B Instruct

262K

€0.18

€0.54

Hermes 4 405B

128K

€0.90

€2.70

GLM 5.1

200K

€1.26

€3.96

GLM 5.2

€1.26

€3.96

Reasoning

CONTEXT

INPUT

OUTPUT

gpt-oss-120b

131K

€0.14

€0.54

DeepSeek V4 Pro 0813

979K

€1.19

€3.56

DeepSeek V4 Pro

€1.58

€3.15

Vision

CONTEXT

INPUT

OUTPUT

MiniCPM V 4.5

32K

€0.59

€1.00

Kimi K3

€2.70

€13.50

Embeddings

CONTEXT

INPUT

OUTPUT

Qwen3 Embedding 8B

41K

€0.01

n/a

We’re here to answer all your questions

Do you really charge no egress fees?

Correct, no egress fees under our fair use policy. Fair use means egress in line with normal workloads. If your egress massively exceeds your stored volume over a sustained period, we’ll reach out to find a fair arrangement.

How does Impossible Cloud compare to other providers?

We offer European digital sovereignty with data protection under a single legal framework. Our vertically integrated technology stack delivers enterprise-grade performance, and published prices with no egress or API fees keep costs predictable.

Where is Impossible Cloud headquartered?

Our headquarters are in Hamburg, Germany, and we operate under strict European data protection law.

Who owns Impossible Cloud?

Impossible Cloud is privately owned by its co-founders, Dr. Kai Wawrzinek and Dr. Christian Kaul, together with leading European venture capital firms and investors.

When was Impossible Cloud founded?

In 2021. Since then we have grown to several thousand customers and a strong partner network across Europe.

What’s the difference between per token, per seat and dedicated?

Per token is a serverless endpoint, so you pay only for the tokens you use. Per seat gives each person on your team one monthly price, from €19. Dedicated reserves GPUs for your model, priced per contract, which costs less once your traffic is steady.

Do you store my prompts or train on them?

No. Your prompts and responses aren’t cached or logged, and your data never trains our models.

Is managed inference available today?

Yes. Pick a model, tell us the workload, and we’ll get you set up.

Which models can I run?

Mistral, DeepSeek, Qwen, GLM, Hermes, Gemma, Nemotron, gpt-oss, Kimi and MiniCPM, plus vision and embedding models, hosted in Europe (EU or UK). Every model and its price is in the full price list. New families are added as they are released. If you need another model, tell us in the form.

Is the API OpenAI-compatible?

Yes. Change the base URL and the key, and the rest of your code stays as it is.

How is pay per token billed?

With separate rates per million input and output tokens. Every model’s price is in the full price list on the managed inference page.

Can I pay per seat instead of per token?

Yes. Per-seat plans start at €19 per user per month, for teams that chat with open models. Talk to sales to set one up.

What if I need guaranteed capacity?

Dedicated reserves GPU capacity for your model, and we still run it. Talk to sales to size it.

Can I create geo-redundant backups?

Yes. Impossible Cloud offers multiple regions across Europe, and you can write your data to each region independently.

What kind of performance can I expect from the service?

An always-hot storage model with consistently low latency and high throughput. You can test it on your own workload with a free trial.

How durable is Impossible Cloud Storage?

It is designed for 99.999999999% (11 nines) durability over a year. Erasure coding distributes data across servers, so disk or server failures don’t lose data. We still recommend versioning, access controls and regular backups.

What integrations and backup solutions do you support?

Most customers use us with Veeam, Acronis, MSP360, Nakivo and Hornetsecurity, with deep or native integrations.

Do you offer Object Lock (WORM) functionality for immutable storage?

Yes. Object Lock prevents data from being changed or deleted for a defined retention period, for compliance and ransomware protection.

Can I migrate my data from another provider to Impossible Cloud?

Yes. Full S3 compatibility makes migration straightforward, and reserved capacity plans include migration support at no extra cost.

Are GPU resources dedicated per customer or shared?

Dedicated. You get full access to NVIDIA systems, with no sharing and no virtualization layer.

How is bare metal GPU usage billed?

Per GPU-hour, on a fixed term from 1 to 60 months, depending on the cluster. Longer terms get a lower rate, and you see the full price before you order.

Who is the contractual partner and how is support handled?

Impossible Cloud is the contractual partner, and support is provided through our service framework.

What security certifications do you have?

Impossible Cloud is ISO/IEC 27001 certified as a company, and our data centers are certified to ISO 27001, SOC 2 and PCI DSS.

How is my data protected at rest and in transit?

Data is encrypted in transit and at rest, using industry-standard algorithms.

How can I manage access to my data?

Identity and access management includes role-based access control and multi-factor authentication.

Run inference in Europe

Managed inference that fits how you work. Three deployment models. One API.

Per token

Per seat

Dedicated

One API for all three. Start per token and move to dedicated without changing your code.

Your data stays yours

Not cached

Not retained

Not for training

Not sure where to start? Talk to an expert

Works with the tools you already use

Get started in your code

Full price list

We’re here to answer all your questions

Start building

Run inference
in Europe

Managed inference that fits how you work.
Three deployment models. One API.

One API for all three.
Start per token and move to dedicated
without changing your code.