Availability
Use health checks and live readiness without treating volatile capacity as a guarantee.
GET https://api.koscompute.com/health is the public health endpoint:
{"status":"ok"}
It intentionally omits backend and scheduler state. HEAD https://api.koscompute.com/ returns 204 and is suitable for a minimal host probe.
API host health does not guarantee that every model or modality is ready. Query GET /v1/models and select entries with is_ready:true. A model can become temporarily unavailable between discovery and admission, so clients must still handle 429 and 503.
Do not continuously poll model discovery from every browser or worker. Cache it for 15–60 seconds, refresh when selecting a model, and use a manual refresh in operational interfaces.
qwen/qwen3.8-27b has no reserved GPU capacity. It runs only while a rental card is free and can stop for a rental or a host GPU check. Loading takes time after a card becomes free; an in-flight request may fail during preemption. No alternate model is served for a Qwen request when capacity is unavailable. Handle 503 with backoff and retry only when your request is safe to repeat.