Sval Compute is an inference API serving models for tasks such as classification, moderation, embedding, and reranking. Zero data retention. Full infrastructure in the EU.
Smaller models are well suited for offloading specialized tasks in agentic systems: they respond faster, cost less per call, and hold up at high call volumes. Sval serves this tier.
Model inference runs on GPU hardware we own and operate in Sweden. The entire serving path, from network entry to model output, stays in Sweden under EU jurisdiction.
Prompts and completions exist in memory only, while being processed. The prefix cache is memory-only with short TTL and per-tenant isolation.
Products
A classification endpoint for content policy checks. Send it text, get back a label from a fixed set of safety categories. Typical usecases are screening user-generated content, checking prompts and model output in an LLM application, and flagging items for human review. The calls are short and cheap, which matters when volume is large.
Currently serving shieldstral-1.0-3b in Live catalog.
Extract fields, transform, summarize, translate, draft, label against your own categories. These tasks can run thousands of times a day, so cost per call matters. A small model does the work well. Optionally, a JSON schema can be sent in the request to constrain the structure of the model output.
Currently serving gemma-4-26b-a4b in Live catalog.
An embedding endpoint that turns text into vectors, so you can match by meaning rather than by keyword. Documents are embedded once and kept on your side, while each query is embedded as it arrives, so the calls run at volume.
Currently serving bge-m3 in Live catalog.
Contact
Sval Compute is a Swedish inference provider, built and operated by Albin Pansell and Simon Riis. API access is currently available on request: write to us with a few lines about what you want to run, and we will get back to you with access details! If there is a model you would like us to host, we are happy to look into it!