Swiss AI Research Platform

Frequently Asked Questions

Common questions about access and use of the Swiss AI Platform

What is the SwissAI Research Platform?

An LLM inference service for the academic community, built as part of the Apertus project. It gives you an OpenAI-compatible API and a Chat interface for models we host ourselves or through our partners. With a CSCS compute grant you can easily launch and serve your own models on CSCS via Slurm.

Which option should I use?

Follow the questions below to find the best fit for your use case. Valid for academic use only:

Yes No Yes No Yes No Yes No Yes No Process or generate huge amounts of data (e.g. in a batch) as fast as possible? Do you have a CSCS grant or project with compute? Individual chat, RAG or agentic usecases? Currently under 15 requests/minute for provider models, slightly higher for platform-served models - click for details. (low request rate) Is the model already hosted here or via a provider? Need a production-grade, always-on endpoint? How to launch your own models - see FAQ below. Use SwissAI Research model launch on Slurm ↓ Compute grant tiers and process - see FAQ below. Apply for a compute grant, or go directly to a provider ↓ Use SwissAI Research need more? see FAQ ↓ How to launch your own models with model launch - see FAQ below. Launch it yourself (needs a grant), or contact us to request ↓ CSCS Inference and the other options - see FAQ below. Use CSCS Inference per-token billing with production autoscaling ↓ Email and Slack - see FAQ below. Contact us - tell us what you need ↓

Why did you make this?

We built this service for research purposes as part of the Apertus project - while building our own models, we also needed a place to test and run them. Opening that platform up lets the wider academic community use it too.

What is different about the SwissAI Research Platform?

We can host models on both Slurm and Kubernetes - a combination that is unique to this platform, made possible by OpenTela routing. That means the same OpenAI-compatible API can serve models from batch-style GPU allocations on CSCS Alps as well as always-on Kubernetes deployments.

What repos make this service?

  • OpenTela - the decentralized compute fabric that routes requests to models wherever they run (we deploy from a swiss-ai fork).
  • serving-api - this website: the API gateway, budgets and the frontend.
  • model-launch - tooling to launch and serve your own models on Slurm.
  • infra - infrastructure and ArgoCD deployment configs (private).

What are providers?

Providers are external inference platforms we have a partnership with. We route requests through our SwissAI Research platform to them, so you can reach their models with the same API key and OpenAI-compatible API you already use here. All current providers are Swiss universities, so access remains academic. Requests to providers are currently rate-limited to 15 requests per minute per user.

I want to use LLMs, what options are available?

(0) SwissAI Research (this website) - hosted models via the OpenAI-compatible API and the Chat app; see the models list.

(1) CSCS Inference - a curated list of always-on models with production-grade autoscaling.

(2) EPFL RCP AIaaS - curated models on the EPFL RCP cluster.

All of these have models that change over time.

I want to run my own LLM, what options are available?

On SwissAI Research:
  • Slurm, via model launch - the standard route, using your CSCS GPU grant.
  • Kubernetes - in special circumstances; please contact us.
See How do I launch my own models? below for the details.

When should I use each?

  • Models already listed here - the default choice for everyday API and Chat use; no setup, no grant needed.
  • Slurm via model launch - when you need high throughput or want to batch-process large amounts of data, or when we don't host the model or configuration you need. Requires a CSCS GPU grant.
  • Kubernetes - special cases that need an always-on, production-style deployment; contact us to discuss.
  • CSCS Inference - when you want a stable, production-grade endpoint for one of their curated models and the rate limits here are too tight.
  • EPFL RCP AIaaS - when you are EPFL-affiliated or their curated model list fits your needs.

What if the provider rate limits are not enough?

You have two options:

(1) Use model launch to launch your own model - we recommend this if you want to process large amounts of data in batch and need high throughput.
(2) Go directly to the provider and request access with them.

Why do you route to providers instead of hosting everything yourself?

Often the same models are already hosted by a provider (for example on CSCS), so hosting them twice would waste GPU resources. Routing also adds resiliency: if CSCS has an outage, traffic can still be served via RCP.

What is the tech stack behind a provider?

It is best to ask the provider directly, as their stack may change over time - see the provider links above for their documentation and contact details.

Can I use this in my project?

Yes, if it's an academic project. Researchers at Swiss universities can sign in and use the platform directly. For projects outside those universities, it needs to be focused on Apertus.

Can I use this for commercial or startup use?

No, this is for non-commercial use only. Please reach out at swissaiserving@ai.ethz.ch, including some details and we will forward your request to the right internal team or industry partners.

Can I launch any model?

Yes, see model launch. This usually requires a CSCS GPU grant. Compute grants come in two tiers: small (<32k GPU hrs) and large (>500k GPU hrs). See the compute grants process.

How do I launch my own models?

Use model launch - a framework-agnostic tool that submits Slurm jobs for distributed inference with vLLM or SGLang on CSCS Alps, for single- or multi-node deployments. It registers your model with OpenTela automatically, so it becomes reachable through this platform's API and appears under your username. You need a CSCS GPU grant (see the previous question). Start with the model launch guide, then the repository's examples for popular models.

Is my institution enabled?

Every account gets an active budget on its first sign-in. If you cannot sign in, or your account shows as inactive, please contact us.

I'm at a non-Swiss university, can I get access?

Yes, but service is prioritised in this order:

(1) ETH and EPFL,
(2) Swiss AI-related institutions,
(3) global academic institutions that focus on Apertus.

I don't have an account at one of the supported universities. How do I get access?

Sign-in offers only the Swiss university accounts we federate with, so there is no way to register yourself otherwise.

Access from outside those universities is reserved for work on Apertus. If your project is Apertus-focused - for example evaluating, fine-tuning or building on the Apertus models - email swissaiserving@ai.ethz.ch with a short description of the project, how it involves Apertus, and who you are working with, or post it in the Swiss AI Serving Slack channel.

Can I get access as a guest or external collaborator?

Only for Apertus-focused work. Guest accounts are added manually by our team, so please contact us with a short description of your Apertus project and who you are working with at a supported institution. We are not able to offer guest access for general LLM use or for projects unrelated to Apertus.

I'm running a hackathon or sprint, can we get access?

Only if it is academic and focused on Apertus. If so, please reach out at swissaiserving@ai.ethz.ch with the dates, expected number of participants and how Apertus features in the event.

I can't access the Chat because I'm not on the network. What can I do?

If you have API access, try running Open WebUI locally: see the Open WebUI guide. Otherwise, we can whitelist your institution or VPN IP range; please contact us.

I would like to collaborate on an academic project, is this possible?

Yes, if the project is centred on Apertus. Our collaboration capacity goes to Apertus activities: evaluating, fine-tuning or building on the Apertus models, and improving the platform that serves them. Researchers at Swiss universities can sign in and use the platform without any of this; for collaborations from outside, Apertus focus is a requirement. Please reach out at swissaiserving@ai.ethz.ch with a short description of the project and how it involves Apertus.

Do you store my prompts and responses?

For the API: no. We do not store prompt or response content. We keep only metadata logs (e.g. userid, input/output token counts, model, timestamp) for capacity planning and abuse prevention.

For the internal-facing Chat app: yes, your conversation history is retained so you can return to past chats. This is important for the Apertus team's development and testing of various model types. Your chats are not extracted and used for training. To request deletion of your chat history, please contact us.

Do you train models on my data?

No. We do not use your prompts or responses to train or fine-tune models. The data used to train Apertus is subject to strict data standards.

How long do you retain data?

API metadata logs (token counts, model, timestamp) are kept indefinitely for operations and abuse prevention. Individual chat history is retained until you delete it via OpenWebUI settings or you delete your account; to do this, please contact us. For specific data retention questions, please contact us.

For the Chat application, where is my data processed and stored?

The platform runs on Swiss infrastructure (CSCS), which is based in Lugano. This includes the entire application, including the web application, API, database and storage.

Can I send personal, confidential, or sensitive data?

Do not send personal, confidential, or otherwise sensitive data (e.g. personally identifiable information). The platform is a research service and is not intended for processing such data. If your project has specific data-protection requirements (FADP/GDPR), please contact us first.

What models are available?

You can list the currently available models on the webpage or via the /v1/models endpoint, or pick from the model selector in the Chat app.

Are there rate limits or usage budgets?

Yes. Each user/institution has a usage budget, and requests may be rate-limited to keep the service fair and available. A low request rate is currently under 15 requests per minute per user for provider-routed models, and slightly higher for models served directly on the SwissAI platform. If you need a higher limit for a project, please contact us.

How do I get or reset my API key?

You can create and reset your key on the API key page. Currently there is only one API key available per user.

Can I use OpenAI-compatible tools and SDKs?

Yes. The API is OpenAI-compatible, so it works with the OpenAI SDKs, LangChain, Open WebUI and similar tools - just point the base URL at our endpoint (.../v1) and use your API key.

How should I cite or refer to the service in a paper, proposal, or other document?

Please refer to the service as Swiss AI Serving and cite the OpenTela project that powers it using the following BibTeX entry: 
@inproceedings{318577,
  author = {Xiaozhe Yao and Youhe Jiang and Ilia Badanin and Qinghao Hu and Robert Matthew Smith and Binhang Yuan and Imanol Schlag and Eiko Yoneki and Ana Klimovic},
  title = {{OpenTela}: Unifying Decentralized Computing Resources for Heterogeneous {LLM} Serving (Operational Systems)},
  booktitle = {20th USENIX Symposium on Operating Systems Design and Implementation (OSDI 26)},
  year = {2026},
  isbn = {978-1-939133-55-7},
  address = {Seattle, WA},
  pages = {1821--1838},
  url = {https://www.usenix.org/conference/osdi26/presentation/yao},
  publisher = {USENIX Association},
  month = jul
}

How do I contact you?

(1) Email - swissaiserving@ai.ethz.ch is the address for all correspondence: access requests, higher limits, data-protection questions, deletion requests and anything else.

(2) Community Slack - the SwissAI Initiative Slack is a good place for community discussion and quick questions (requires a SwissAI Initiative account).