Mycelis
Mycelis
Model hosting without GPU know-how
You do not rent a GPU — you pick a model. Mycelis starts it on a suitable GPU, sizes it to your number of users and only bills while it runs. Switch it off overnight, switch it on in the morning.
No base fee · No minimum term · Pause any time
What Mycelis takes care of
Everything behind a single address
api.mycelis.ai/proxy/v1
The common misunderstanding
Mycelis has no server catalog. You tell us which model you want to use — hardware, runtime and operations come with it.
Classic GPU hosting
With Mycelis
How it works
No server setup, no GPU selection — from an empty workspace to a running model.
0@i
Pick an open-source model from the catalog and say how many people will work with it at the same time — or start a ready-made app such as OpenWebUI.
0@i
We choose a suitable GPU, start the runtime and set up networking and HTTPS.
0@i
You get an OpenAI-compatible endpoint and a chat interface. When the model runs is up to you: by click or on a schedule.
Simplified demo. How long a real deployment takes depends on the model — the weights have to be loaded onto the GPU first.
Cost control
Your team does not work at night? Then the GPU does not have to run either. You decide when your model is on — and Mycelis makes sure it is off the rest of the time.
Switch on and off with one click
Stop the model overnight or at the weekend. While it is stopped, no GPU runtime accrues. Start it again the next day — the deployment and its settings stay in place.
A schedule for your working hours
Set a daily time window. Mycelis starts the model when the window opens and stops it when it closes — no one has to remember.
Stop on inactivity
Optional: the model stops by itself after a period without requests that you define.
Wake on request
Optional: a stopped model starts automatically when the first request arrives (allow for the model's load time).
Limits against surprises
A daily runtime limit per deployment and a monthly budget per workspace. When a limit is reached, the deployment is stopped.
Runtime across 24 hours (example)
58 % less runtime in this example
Simplified example. Actual costs depend on the GPU that is chosen.
Scaling
You say how many people work with the model at the same time. Mycelis works out the memory it needs and picks a suitable GPU size — no more than you need.
Think in people, not gigabytes
Number of concurrent users instead of a GPU catalog and VRAM tables.
Adjust any time
When the team grows, raise the number — the deployment is set up again on a suitable GPU.
Training works the same way
Fine-tuning runs automatically get a GPU with enough memory.
Example: a 32B model
Simplified illustration. The actual choice depends on the model, quantization and availability.
More than hosting
Everything below is already there when your use case asks for it. Start with one model, add a layer at a time.
Combine a model with a system prompt, tools, knowledge bases and MCP servers — as rule-based routing or a visual workflow. Callable like any other model.
Train an adapter (LoRA) on your own data and deploy the result like any other model. Billed by compute time.
Upload documents, code or PDFs. Relevant passages are attached to the request automatically.
Give your models access to external tools and APIs from a curated catalog, without touching client code.
Route by token count, keywords or tool use — or let the automatic router pick the right model for each request.
Detects personal data in requests and keeps them off external providers if your policy says so.
Ready-made apps
Popular tools that work with your models — hosted, connected and billed by the hour.
A chat interface for the whole team, connected to your hosted models. One click, from $0.05 per hour.
Upload documents and let models answer from your own material. $0.05 per hour plus $0.01 per GB and hour.
Workflow automation with your own models as building blocks.
Node-based image workflows on your own GPU.
For very different teams
A regulated enterprise, a business without an IT department, a developer team — the same platform, set up differently.
Every department works in its own workspace with its own roles, permissions and monthly budget. Requests containing sensitive data can be kept on models you host yourself, and every request is recorded — who sent it, which model answered, what it cost.
You do not want to run inference — you want your support inbox triaged and your product texts written. Host a model with one click, give the team a chat interface and switch it off when nobody is working.
Mycelis speaks the OpenAI API, so Cursor, Continue, Claude Code and your own SDK calls keep working. Routing rules decide per request which model answers — a small open-source model for routine completions, a frontier model when the task actually calls for it.
Automatic routing
Every request is checked against your routing rules. Mycelis only escalates to an expensive frontier model when complexity, context length or tool use actually call for it.
You define when a request may escalate to a larger model — and every decision stays traceable.
One endpoint, visible decisions
Your application always calls the same URL. Which model answered is recorded for every request — no black box.
Rules you control
Route by token count, keywords, tool calls or regular expressions — or let the automatic router decide.
Per request
The decision is made for each individual request, not once per application. A workload that is mostly routine only pays frontier prices for the part that really needs it.
Data protection by design
Mycelis inspects every request before it leaves the gateway. You decide what counts as sensitive — and anything that matches goes to a model you host yourself instead of an external provider.
Automatic filter for personal data
Two levels: flag the request and notify an admin, or keep it off external providers altogether.
Your own models for sensitive work
Open-source models on a GPU deployment that belongs to your workspace. Commercial providers never see these requests.
BYOK or managed keys
Use your own OpenAI, Anthropic or Google keys, or let Mycelis manage them. Switching is a setting, not a migration.
Roles, permissions and budgets
Every workspace has roles, permissions and a monthly budget — configured in the dashboard, not through a support ticket.
Simplified illustration of the routing decision
For developers
Mycelis speaks the OpenAI protocol. In most projects, switching over means changing two lines: the base URL and the key.
from openai import OpenAI client = OpenAI( base_url="https://api.mycelis.ai/proxy/v1", # these two lines are the only change api_key="pat_…" ) response = client.chat.completions.create( model="qwen3-32b-instruct", # any model you activated in the workspace messages=[{"role": "user", "content": "Hello"}] )
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.mycelis.ai/proxy/v1", // these two lines are the only change
apiKey: process.env.MYCELIS_API_KEY
});
const response = await client.chat.completions.create({
model: "qwen3-32b-instruct", // any model you activated in the workspace
messages: [{ role: "user", content: "Hello" }]
});
# these two lines are the only change curl https://api.mycelis.ai/proxy/v1/chat/completions \ -H "Authorization: Bearer pat_…" \ -H "Content-Type: application/json" \ -d '{ "model": "qwen3-32b-instruct", "messages": [{"role": "user", "content": "Hello"}] }'
What happens to a request
Pricing
No platform fee, no seats, no minimum term. Every component is billed separately, by actual usage.
Model hosting on GPU
Billed at the hourly rate of the GPU that is chosen, to the minute — only while the model runs. The rate depends on the GPU size.
Commercial models
Billed per token. Bring your own key and pay the provider directly — or use a key managed by Mycelis.
OpenWebUI chat interface
$0.05 per hour while it runs.
Knowledge bases
$0.05 per hour plus $0.01 per GB and hour.
Training and fine-tuning
Billed by actual compute time. No separate package, no minimum term.
Calculate your real cost
For a self-hosted ~32B-parameter model. Prices below are live GPU rates from our catalog, not estimates.
Per hour
$0.77
Per month, always on
$565
Per month, 8h/day
$136
Illustrative sizing for a ~32B-parameter model (base VRAM + 0.5 GB per concurrent user). GPU prices are live; actual cost also depends on the specific model you deploy and how long it runs — you're billed only while it's running.
Common questions
No. You choose a model. Mycelis picks a suitable GPU, starts the runtime and sets up networking and HTTPS. You never work with GPU types or drivers.
You state how many people should work with the model at the same time. From that and the size of the model, Mycelis works out the memory required and picks a GPU that fits. If the team grows, you raise the number and the deployment is set up again on a suitable GPU.
The deployment is stopped and no GPU runtime accrues while it is stopped. The deployment and its settings stay in place. The next day you start it again with one click — or it starts on its own if you use a schedule or wake-on-request.
Yes. You can set a daily time window in which the model runs, a one-off window, an automatic stop after a period without requests, a daily runtime limit and wake-on-request. On top of that there is a monthly budget per workspace.
That depends on the model: the weights have to be loaded onto the GPU first, and the larger the model, the longer it takes. You can follow the progress live in the dashboard.
Pure pay-as-you-go. A hosted model is billed at the hourly rate of its GPU, to the minute, while it runs. Commercial models are billed per token, and apps such as OpenWebUI per hour. No base fee.
A deployment can only be started with sufficient balance. In addition, you can set a monthly budget per workspace; when it is exceeded, running deployments are stopped.
Build agents, train and fine-tune your own models, connect knowledge bases (RAG) and MCP tools, route requests automatically between models and start ready-made apps such as OpenWebUI. All of it is reachable through the same OpenAI-compatible endpoint.
Yes. Train an adapter (LoRA) on your own data with Mycelis and deploy the result like any other model in your workspace.
A hosted open-source model runs on a deployment that belongs to your workspace; requests to it do not go to commercial providers. An optional filter detects personal data and can keep such requests off external providers. Details are in our privacy policy.
Yes. Commercial models can be activated with your own key (BYOK) or a key managed by Mycelis. The endpoint is OpenAI-compatible: you only change the base URL and the API key — request and response format stay the same, streaming included.
No. There is no base fee, no seat price and no minimum term. You only pay for the components you actually use.
Both are on our roadmap. We do not promise a date yet — the apps that are available today are marked as such on this page.
Ready to get started?
Pick a model, let Mycelis host it and switch it off when nobody needs it. No base fee, no minimum term.