Build on Nebula.

Introduction

Nebula is an inference cloud for frontier AI. You deploy any open model, then serve billions of tokens over a single API — with autoscaling, 8ms cold starts and per-token billing.

All requests are made over HTTPS against https://api.nebula.run. Responses are JSON and OpenAI-compatible.

Quickstart

$ curl https://api.nebula.run/v1/completions \
    -H "Authorization: Bearer nk_..." \
    -d '{"model":"llama-3.1-70b","prompt":"Hello,"}'

API reference

The core resource is a Model. Deploy one, then call it over the completions or embeddings endpoints.

Models

Nebula hosts 200+ open models — Llama, Mistral, Qwen, Gemma and more — plus any fine-tuned checkpoint you upload. Models are kept warm on standby so cold starts stay in single-digit milliseconds.

Scaling

Autoscaling is built in. Nebula scales your fleet from zero to thousands of GPUs as traffic changes, so you never provision for peak and never drop a request.

Fine-tuning

Upload a dataset, start a fine-tune job, and serve the resulting checkpoint through the same API. Your data is never used to train shared models.

Observability

Every request reports token usage, latency and cost. Dashboards, alerts and per-model breakdowns are built in.

Security & compliance

We are SOC 2 Type II audited. Encryption everywhere, VPC deployment options, and field-level access controls.

Guides

• Inference is the new bottleneck
• Cold starts, solved