Hosted in Switzerland · Open-weight models · OpenAI-compatible
Open-weight models on a GPU of your own, behind an API that the OpenAI SDKs already speak. Prompts and answers are not stored, and all our servers are in data centres in Switzerland.
# Python, official OpenAI SDK
from openai import OpenAI
client = OpenAI(
base_url="https://api.llmaas.ch/v1",
api_key="llmaas_…",
)
r = client.chat.completions.create(
model="gemma-4-26b-a4b",
messages=[{"role": "user",
"content": "Summarise this…"}],
stream=True,
)Solo runs Gemma 4 26B-A4B or Qwen3.5-9B on an NVIDIA L4; Team and Scale run larger models on bigger cards. Your own fine-tune can run too, if it fits the GPU.
A virtual machine with its own NVIDIA GPU in a data centre in Switzerland, running vLLM. It serves your account and nobody else.
An OpenAI-compatible base URL, API keys for your applications, and the web chat at chat.llmaas.ch, with file upload and web search.
Requests go through our gateway to your box in Switzerland and are answered there. Prompts and answers are not stored and not used to train any model; our logs record token counts, not content. A file attached in the chat is held in memory for at most an hour.
The GPU serves only your requests, so its speed depends on your own load alone. You pay a monthly price for the box, not per token. The API accepts 20 requests a second per client address, with bursts up to 100.
The OpenAI SDKs, LangChain, OpenCode, Cursor and any other tool that lets you set the base URL of an OpenAI-compatible API. Streaming, tool calling, thinking mode and image input are supported.
The models, from Qwen and Google, are Apache 2.0 and their weights are public. If you leave, they run on your own hardware.
At chat.llmaas.ch, signed in with your LLMaaS account: document upload, a thinking switch and a meter of the context used. Conversations stay in your browser. Web search can be switched off; when it is on, its queries go to DuckDuckGo, outside Switzerland, and our server opens the pages the model picks.
LLMaaS is operated from Prilly, in the canton of Vaud, and its servers are in data centres in Switzerland. Support answers by email in English, French or German. The contract is under Swiss law and written in English; invoices are in Swiss francs.
Two models we recommend. Both read images, both are Apache 2.0, and both handle up to 256K tokens of context; the context each plan serves is listed under Pricing. Gemma 4 26B-A4B is the Solo model; Qwen3.6-35B-A3B needs the Team GPU. Solo can also run Qwen3.5-9B, a smaller Qwen that leaves room for a longer context.
Alibaba's sparse mixture-of-experts model, made for agentic coding, function calling and structured output. The one we suggest for applications that call tools.
Google's mixture-of-experts model. The fastest of the models we run on the L4, and the one we suggest for documents, screenshots and diagrams.
On the Team plan, Qwen3.6-35B-A3B, Qwen3.6-27B and Gemma 4 31B run in FP8; 100B-class models run on the Scale plan. Any other model vLLM can serve is deployed on request, if it fits the plan's GPU and its licence allows it.
A personal assistant that runs around the clock on a small machine of its own, using the LLMaaS model, your GPU box or another provider. You write to it where you already are, it uses your tools, and it remembers.
Its own web chat, and the channels you connect: Discord, Telegram, Slack, Matrix, email. It keeps one memory across all of them.
It runs commands, reads and writes files, reads web pages and works with git on its own machine, so it can carry a task through to the result.
Add MCP servers from its settings: your calendar, your code host, your notes. The agent then reads and acts there for you.
Ask for a summary every morning or a check every hour. It runs the task by itself and writes to you when there is something to say.
What it learns about you and your work stays between conversations, on its own disk. Download that memory whenever you like.
Its disk is encrypted, and you choose who can unlock it: you alone, or LLMaaS too, so that it starts again by itself after a reboot. The machine is in Switzerland and has no SSH access: it is reached only through the LLMaaS gateway.
CHF 19 a month, on its own or with a GPU plan. On its own it includes 100 requests a day on the LLMaaS model; with a GPU plan it uses your box without a daily limit. It can also run on your own key with another provider.
Each plan is a VM with its own GPU in Switzerland. Every plan includes the chat, API keys and unlimited tokens. Invoiced monthly; you can cancel for the end of any month.
GPU plans are on request. Demand for GPUs is high and capacity in Switzerland is limited, so we confirm a GPU is free for you before we open your VM.
from CHF 399/month
NVIDIA L4 · 24 GB
from CHF 799/month
NVIDIA L40S · 48 GB
from CHF 1 999/month
NVIDIA A100 · 80 GB
CHF 19/month · on its own, or with any plan
An always-on assistant on a small VM of its own, using the LLMaaS model, your box or another provider: chat channels (Telegram, Slack, email…), shell and file tools, scheduled tasks, and a memory it keeps between conversations.
Starting prices; no VAT is charged at present. A VM is billed by the hour, at the monthly price divided by 730, from the moment it is yours until it is paused or deleted: a 30-day month costs a little less than the monthly price, a 31-day month a little more.
Yes, with one exception in the chat. Your model runs on a virtual machine that serves your account alone, in a data centre in Switzerland. Requests travel encrypted from your application through our gateway to that machine and back, and nothing of them is stored once the answer is sent. The exception: when web search is on in the chat, its queries go to DuckDuckGo, outside Switzerland, and our server opens the pages the model chooses. The API never searches the web.
Yes, if it uses the OpenAI chat completions API: set the base URL and the key. That covers the official SDKs, LangChain, LlamaIndex, OpenCode, Cursor, Continue and most agent frameworks. Streaming, tool calling and image input are supported.
Qwen3.6 if your application calls tools or writes code; it needs the Team plan. Gemma 4 if you want the fastest chat or read many documents and images. We change the model on your box when you ask, if the new one fits the plan's GPU; it takes a restart.
On the Solo plan a single user sees 30 to 45 tokens per second, and the GPU serves around 15 chats at once before individual speed drops. The Team plan roughly triples that. Long documents take a few seconds to read the first time and are cached afterwards.
Yes, if vLLM can serve it, it fits your plan's GPU and its licence allows it; that includes your own fine-tunes. Tell us the Hugging Face repository and we tell you which plan it needs.
LLMaaS is operated from Prilly, in the canton of Vaud. Support answers by email at hello@llmaas.ch on business days, by the next business day at the latest. The contract is under Swiss law, and invoices are in Swiss francs.
Pick a plan and tell us your use case and your team size. We confirm a GPU is available for you in Switzerland, then send you your base URL and key.
Request accessOr write to us: hello@llmaas.ch