No pre-provisioned instances — scale up on traffic, scale to zero when it leaves. Billed per token and second, cold start in seconds.
Focus on business, we handle infrastructure.
Pulled up on request, no always-on.
Elastic with concurrency, no drop at peaks.
Per token/second; zero cost when idle.
Built-in vLLM and mainstream open models.
Per token and runtime, finely metered; no charge when no request.
vLLM and mainstream open models, hot-swappable.
Seconds — fits bursty traffic.
Steady high load → reserved; spiky → Serverless.