Serverless Inference Closed beta

OpenAI-compatible inference.
Hosted in Europe.

Run Llama, Mistral, Qwen and more through a vLLM-powered, OpenAI-compatible API. Pay per token. Your data stays in the EU.

Talk to an engineer

20 min, no slides. Fast track into the beta.

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://iris.lyceum.technology/v1",
    api_key="<your api key>",  # from API Keys page
    default_headers={"User-Agent": ""},
)

response = client.chat.completions.create(
    model="meta-llama/Llama-3.1-70B-Instruct",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=256,
)

print(response.choices[0].message.content)

Migrate from OpenAI by changing base_url and api_key. No other code changes.

Backed by

You need an LLM API. Your compliance team said no to OpenAI.

Compliance is a hard stop

Pharma, finance, defense, public sector: sending data to US-hosted APIs is a regulatory block.

Self-hosting is a full-time job

vLLM, Kubernetes, auto-scaling, GPU cooling, on-call. That is a team you do not have.

Hyperscaler GPUs are expensive

Azure H100 at ~19 EUR/hour. AWS makes you block-reserve. GCP auto-scaling does not work.

Three steps. Under 60 seconds.

Get a beta key

pip install openai

Book a call or join the waitlist. Beta users get a key at dashboard.lyceum.technology.

Change one line

client = OpenAI(
  base_url="https://iris.lyceum.technology/v1",
  api_key=<your api key>,
)

Point your existing OpenAI SDK at Lyceum.

Call any model

response = client.chat.completions.create(
  model="meta-llama/Llama-3.1-70B-Instruct",
  messages=[...],
)

Any vLLM-compatible Hugging Face model. EU data centers.

Talk to an engineer about migrating

What you get

base_url="api.openai.com"

base_url="iris.lyceum.technology"

OpenAI-compatible API

Drop-in replacement. Same SDK, same streaming.

EU-WEST-1 Frankfurt, DE

EU data residency by default

All inference runs in European data centers.

$0.02

Pay per token

No minimums, no subscriptions. Published rates.

13 models

vLLM-powered

Deploy any Hugging Face model that vLLM supports.

No infrastructure

No Docker, no Kubernetes. One API endpoint.

Auto-scaling

Scale from zero to thousands of requests.

vLLM-compatible models

Deploy any Hugging Face model that vLLM supports. Here are some popular ones.

Model	Hugging Face ID	Params	Type
Llama 3.1 70B	meta-llama/Llama-3.1-70B-Instruct	70B	text
Llama 3.1 8B	meta-llama/Llama-3.1-8B-Instruct	8B	text
Llama 3.2 3B	meta-llama/Llama-3.2-3B-Instruct	3B	text
Qwen 2.5 72B	Qwen/Qwen2.5-72B-Instruct	72B	text
Qwen 2.5 7B	Qwen/Qwen2.5-7B-Instruct	7B	text
Mistral 7B	mistralai/Mistral-7B-Instruct-v0.3	7B	text
Mixtral 8x7B	mistralai/Mixtral-8x7B-Instruct-v0.1	8x7B	text
Qwen 2.5 Coder 32B	Qwen/Qwen2.5-Coder-32B-Instruct	32B	code
Qwen 2.5 Coder 7B	Qwen/Qwen2.5-Coder-7B-Instruct	7B	code
LLaVA 1.6 7B	llava-hf/llava-v1.6-mistral-7b-hf	7B	multimodal
Whisper Large V3	openai/whisper-large-v3	1.5B	audio
Whisper Large V3 Turbo	openai/whisper-large-v3-turbo	809M	audio
BGE Large EN	BAAI/bge-large-en-v1.5	335M	embedding

Any model supported by vLLM can be deployed. See the full vLLM compatibility list.

How Lyceum compares

	Lyceum	OpenAI	Together AI	Self-hosted
Data residency	EU	US	US	Your choice
GDPR-compliant	Yes	Contractual	Contractual	Your responsibility
OpenAI-compatible	Yes	Yes	Yes	Build it
Open-source models	Yes	No	Yes	Whatever you host
Infrastructure work	None	None	None	Full stack
Pay per token	Yes	Yes	Yes	n/a

FAQ

It says closed beta. What does that mean in practice?

Serverless Inference is available only to customers we have onboarded directly. There are two ways in: a 20-minute call with one of our engineers (fastest: you walk away with a beta key and a worked migration example), or the waitlist (we invite in batches as capacity opens).

Is Serverless Inference really OpenAI-compatible?

Yes. Change base_url and api_key in the OpenAI SDK. Chat completions, streaming, and tool use work as expected. Function calling and vision models are supported on models that support them natively.

Where does my data go?

All inference requests are served from European data centers. Your prompts and completions do not transit US infrastructure.

How is billing calculated?

Per token. Published rates for all models. No minimums, no subscription. Beta participants receive free credits discussed on the onboarding call.

Can I fine-tune a model on Lyceum and serve it here?

For custom or fine-tuned models, use Dedicated Inference (a dedicated GPU machine with your Docker image, OpenAI-compatible endpoint). Serverless is for the pre-hosted open-source catalog.

Skip the queue. Talk to us.

A 20-minute call with one of our engineers is the fast track into the beta. Bring a use case, we will show you how to port it from OpenAI. No slides.

Book a 20-min call

OpenAI-compatible inference. Hosted in Europe.