Inference
Open-weight LLMs served with vLLM on CUDA, starting with Qwen3.8-27B. Starting card: NVIDIA RTX PRO 6000, 96 GB, rented.
Inference platform · NVIDIA GPUs
Galerna is building an inference platform for open-weight models: one OpenAI-compatible API on NVIDIA GPUs, per-token access through OpenRouter, and dedicated endpoints for teams that need capacity of their own.
01 · Models
A short list, served properly: every model runs at the precision we publish, and we never quietly serve a lower-precision version.
02 · Products
Three ways to use the same models, from a first test to steady production traffic on GPUs reserved for you.
Galerna as an inference provider on OpenRouter, starting with Qwen3.8-27B. Pay per token with your existing OpenRouter account, with no contract with us needed.
Pay per token · prices set at listing
OpenAI-compatible API keys from us, with a contract, an invoice and a data processing agreement.
A model on NVIDIA GPUs reserved for you and shared with no one, for steady production traffic.
About €2,500 per month per endpoint · planned price
Tokens through OpenRouter, reserved capacity direct.Nothing is live yet. Planned prices, for information only; binding terms only in a signed contract.
03 · How it works
OpenAI/Ollama-compatible gateway, prefill and decode metering, and verification that fails closed. 84 automated tests; end-to-end smoke test against a simulated vLLM backend.
First NVIDIA GPUs online, then OpenRouter’s provider process: application, review, an intro call, security questionnaire and terms, go-live. Billing and sign-up for direct customers.
04 · NVIDIA in our stack
We start on rented NVIDIA GPUs and add our own servers as reserved capacity sells.
Open-weight LLMs served with vLLM on CUDA, starting with Qwen3.8-27B. Starting card: NVIDIA RTX PRO 6000, 96 GB, rented.
Our own NVIDIA servers once 60% of their hours are sold, growing into a small fleet in a colocation facility in Spain.
Built in: precision checks before a model is listed. Next: hardware-rooted proof with NVIDIA confidential computing on Blackwell-class GPUs.
05 · Where we are
No GPUs running, no customers and no OpenRouter listing yet. We rent GPUs first and buy our own once their capacity is already sold.
06 · Company
Galerna is a commercial brand of Unisphere Overseas S.L.
Madrid, Spain (EU)
Inference for open-weight models: OpenRouter, direct API, dedicated endpoints
Developers and companies building on open models
Pre-seed. API gateway built; first GPUs and OpenRouter listing next
Generative AI · AI inference
Contact
We’re lining up our first pilot customers for the direct API and dedicated endpoints. Tell us which models you run and at what volume, and we’ll tell you when we can serve them.
Emailinfo@galerna.si