Model Hosting
Open-weight models, hosted with zero data retention.
Metrovolo operates dedicated inference infrastructure for open-weight models used in agentic workloads. Requests are processed in memory and never stored. Nothing is used for training.
What we run
Data policy
Zero retention. Request metadata only. No training on your content.
Regions
EU and US datacenters. The serving region for each model is listed below as it goes live, and EU-hosted models are eligible for in-region routing.
Serving
Dedicated Blackwell-class GPUs, FP8 weights, speculative decoding, prefix caching.
Interface
OpenAI-compatible chat completions with streaming and usage reporting.
Models
Launch model: Step 5 Preview (StepFun), currently in validation. Serving region: to be published at launch. Additional models are added as they pass our serving and quality checks. Live availability, speed, and uptime for each model are published on its OpenRouter page and at status.metrovolo.com.
For platforms and direct customers
We serve requests from routing platforms such as OpenRouter and from direct API customers. Platform users are billed by their platform; direct customers are invoiced monthly per token.
Questions: info@metrovolo.com