Agentic AI, powered from France.
AI agents plan, call tools and chain dozens of model calls to finish a task. Run agentic AI on open-source models through an OpenAI-compatible API or your own GPU endpoints, on our own NVIDIA servers in Paris and Lyon.
OpenAI-compatible · Open-source models · Data stays in France
What is agentic AI?
Agentic AI refers to systems where a language model does more than reply to a prompt: it breaks a goal into steps, decides which tool to call, reads the result and continues until the task is done. An AI agent can search internal documents, query a database, write and run code or update a ticket, with each step driven by the model.
This changes the infrastructure profile. One task can trigger dozens of chained calls, each carrying a long context of instructions, tool results and history. Agentic workflows therefore depend on an AI agent infrastructure with three things: models that reason over long contexts, a cost model that stays predictable as calls multiply, and control over where your data and tool outputs are processed.
Chatbot
AI agent
Input
A single prompt
A goal to reach
Model calls
One per answer
Many, chained in a loop
Context
Short conversation
Instructions, tool results and history
Output
Text
Actions through tools and APIs
What it takes to run AI agents in production.
Agents that work in a demo often struggle at scale. Four constraints come up for every team that builds AI agents.
Long contexts
Every step resends the context
Each call carries instructions, tool definitions and previous results. Agents need models that handle long contexts, and GPU capacity to process them without slowing the loop.
Cost
Costs grow with every step
An agent multiplies model calls. Per-token pricing suits prototypes, and cached input tokens are billed at a lower rate when agents resend the same context. On Serverless GPU you pay per second of GPU time, with no per-token charge.
Data residency
Agents touch your internal data
Agents read documents, tickets and databases. Requests to the Models API and Serverless GPU are processed on our own servers in Paris and Lyon, and your prompts are never used to train models.
No lock-in
Open models, standard API
Open-source models such as Kimi, GLM, DeepSeek, Qwen and MiniMax are widely used for agentic workloads. Reach them through an OpenAI-compatible API and keep the freedom to switch.
AI agent infrastructure, at every stage.
Prototype on a ready-made API, move to your own endpoint when agent traffic grows, and reserve dedicated GPUs for steady, high-volume workloads. Same console, same account.
Per token
Models API
Call open-source models used for agentic workloads through one OpenAI-compatible endpoint. Agent frameworks that accept a custom OpenAI-compatible base URL can point to it.
Best for: prototyping agents, variable traffic, comparing models
Explore the Models API →Per second
Serverless GPU
Deploy a catalog model or any vLLM-compatible Hugging Face model, including your fine-tuned agent model, on the GPU you choose, with scale-to-zero.
Best for: long contexts, heavy agent traffic, custom models
Explore Serverless GPU →Dedicated
Cloud GPU & GPU Clusters
Run your own serving stack on NVIDIA GPU instances hosted in Europe, or reserve a dedicated cluster on our own servers in Paris and Lyon for very high, steady agent volume.
Best for: high, predictable volume and full control
Explore Cloud GPU →Recommended GPUs
On Serverless GPU and reserved GPU Clusters, our own servers in Paris and Lyon: RTX PRO 6000 Blackwell for cost-efficient agents · B300 for the largest models and long contexts
Popular models for agents
Kimi, GLM, DeepSeek, Qwen, MiniMax, GPT-OSS and more. Building coding agents? See AI for coding.
Build AI agents, from first loop to scale.
Test
Compare models on your own agent prompts in the console playground.
Connect
Set the Models API as the OpenAI-compatible base URL in your agent code or framework.
Customize
Deploy a fine-tuned or custom agent model on Serverless GPU, billed per second.
Scale
Move steady agent traffic to Cloud GPU instances or a reserved GPU cluster.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://console.arkanecloud.com/api/v1",
api_key=os.environ["ARKANE_CLOUD_API_KEY"],
)
response = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4.1-Flash",
messages=[{"role": "user", "content": "Hello!"}],
)
Which agentic AI option fits your workload?
A quick guide based on how your agents run, from first prototype to production.
Workload
Recommended
Why
Prototype an agent or compare models
Nothing to deploy, you only pay for the tokens used
Long-context, high-volume agents
Billed per second of GPU time, with no per-token charge on long prompts
Custom or fine-tuned agent model
Deploy any vLLM-compatible Hugging Face model, scale to zero
Agents on sensitive internal data
Any option
The Models API, Serverless GPU and GPU Clusters run in France, Cloud GPU in Europe, and prompts are never used to train models
Very high, steady volume
Dedicated capacity and predictable costs
Coding agents in your IDE
Free extension for VS Code chat and agent mode
Questions about agentic AI
What is agentic AI?
Agentic AI describes AI systems that pursue a goal with a degree of autonomy: a language model plans the steps, calls tools or APIs, reads the results and decides what to do next until the task is complete.
What is the difference between agentic AI and generative AI?
Generative AI produces content in response to a prompt. Agentic AI uses generative models as the reasoning engine of a loop that takes actions, such as searching documents, calling APIs or writing files, to reach a goal.
How do AI agents work?
The application sends the model a goal, instructions and a list of available tools. The model chooses an action, the application executes it and returns the result, and the loop repeats. Each step is a new model call with a growing context.
How do I build an AI agent on Arkane Cloud?
Create an API key in the console, then point your agent code or framework to the Models API through its OpenAI-compatible base URL. When you need a custom model or more capacity, deploy it on Serverless GPU or Cloud GPU.
Do Arkane Cloud models support tool calling?
Yes, on models that support it. The Models API accepts the OpenAI-compatible tools and tool_choice parameters, so your agent can declare functions and receive tool calls, and structured output works the same way through response_format. Support depends on the selected model: check it in the model catalog.
Which agent frameworks work with Arkane Cloud?
Any framework that accepts an OpenAI-compatible base URL and an API key, for example LangChain, CrewAI or AutoGen. Set the base URL to https://console.arkanecloud.com/api/v1, add your Arkane Cloud API key and pick a model ID from the catalog, such as deepseek-ai/DeepSeek-V4.1-Flash.
Which open-source LLMs can I use for AI agents?
The Models API catalog includes open-source models widely used for agentic workloads, such as Kimi, GLM, DeepSeek, Qwen and MiniMax. You can also deploy any vLLM-compatible Hugging Face model on Serverless GPU.
Where is my agents' data processed?
The Models API, Serverless GPU and GPU Clusters run on our own NVIDIA GPU servers in colocation datacenters in Paris and Lyon, France, and Cloud GPU instances are hosted in Europe. Prompts and outputs, including the tool results your agents send to the model, are never used to train models.
How is agentic AI usage billed?
The Models API is billed per million input, output and cached tokens. Serverless GPU and Cloud GPU are billed per second of GPU time, with no per-token charge, which suits agents that send long prompts. GPU Clusters are reserved for 2 to 5 years.
Other use cases
Build your agents in France, today.
Test models in the playground, connect your agent with an API key and scale at your own pace.
Free playground: 10 free requests · No commitment