Skip to main content

Serverless API quickstart

ModelPilot has two paths: dedicated GPU pods for custom ComfyUI workflows, and serverless per-request API for popular open-weight models. This page covers the serverless path. If you have a custom workflow with custom nodes or LoRAs, you want dedicated instead.

When to use serverless (and when not to)

Use serverless when all of these are true:

  • The model you need is in the catalog below
  • You don't need custom ComfyUI nodes, your own trained LoRAs, or multi-step workflows
  • Your traffic is bursty or low-volume (idle costs hurt you)

Use dedicated when:

  • You have a working ComfyUI workflow.json you want to host as-is
  • You need custom nodes, custom LoRAs, or a model not in the serverless catalog
  • You run sustained traffic (8+ hours/day on the same model)

Honest note: for plain “give me Flux Schnell on demand” without ComfyUI, Replicate is often a better choice — broader catalog, faster warm cold starts, lower per-request floor on schnell-class models. See the honest comparison. We exist for users who want both serverless AND custom-workflow dedicated from one provider.

30-second quickstart

Get an API key from your dashboard . Add prepaid credit before the first billable request. Then:

curl -X POST https://modelpilot.ai/api/v1/generate/image \
  -H "Authorization: Bearer mp_live_..." \
  -H "Content-Type: application/json" \
  -d '{"model": "flux-schnell", "prompt": "a red apple on a white table"}'

Two response shapes:

  • Warm hit: JSON with images array, sub-second to a few seconds
  • Cold start: { status: "processing", poll_url: "..." } — poll the URL until it returns COMPLETED

What's available — real models, real prices

Prices are charged only on successful completion — failed or throttled jobs are not billed.

Image

ModelIDPriceFirst request
Flux Schnellflux-schnell$0.008/image~30s cold, sub-second warm
Flux Devflux-dev$0.015/image~30s cold, sub-second warm
Stable Diffusion XLsdxl$0.005/image~25s cold, sub-second warm
Z-Image Turbozimage$0.008/image~25s cold, sub-second warm

Audio (TTS)

ModelIDNotes
Kokoro TTSkokoro-serverless82M params, 50+ voices, 7 languages
Chatterbox Turbochatterbox-serverless350M, expressive, voice cloning

Text (chat completions, OpenAI-compatible)

Qwen3 4B/8B/32B, DeepSeek R1 8B/14B, GLM-Z1 9B/32B. Per-request pricing $0.01-$0.04 depending on size. First request after a long idle is slower (2-4 min) since the model has to download to a fresh worker; warm cold starts are ~25-35s. Full payload format in the full API reference.

Video

Wan 2.2 T2V is in the model config but the serverless image is not yet built — coming soon. For video today, use the dedicated path via Photo Animator (Wan 2.2 5B image-to-video on a hosted A6000).

Cold start, idle, and what gets charged

  • First request, completely cold: 25-30s for image, 2-4 min for text models that need to download weights.
  • Warm requests: sub-second to a few seconds depending on model and prompt complexity.
  • Idle timeout: 15 minutes. After that, the next request hits a warm cold start (~25-35s) since the worker may be recycled.
  • Billing: charged on COMPLETED. Throttled jobs that never run, or jobs that error out, are not billed.
  • GPU allocation: we have a 12-GPU fallback list per model (24-48GB secure + community cloud). RunPod schedules your job on whichever has capacity.

Cost examples

10 Flux Schnell images/day: 10 × $0.008 = $0.08/day, ~$2.40/month
1,000 Flux Dev images/day: 1,000 × $0.015 = $15/day on serverless. Compare to ~$15-20/day on a dedicated A6000 running 8 hrs — at this volume, dedicated wins on consistency, serverless wins on flexibility.
1 hour of Kokoro TTS at ~5 requests/min: ~300 requests, billed per request — see /pricing for current TTS per-request rate.

Limitations to know

  • No custom workflows on serverless yet. If your workflow has custom nodes or trained LoRAs, use dedicated.
  • Catalog is intentionally narrow. ~10 popular open-weight models, not thousands. If you need a specific fine-tune that's not listed, dedicated is the path.
  • Cold starts are real on first hit. If your app cannot tolerate ~30s first-request latency, keep a warm dedicated pod instead.
  • Wan T2V serverless is not live yet — the image hasn't been built. Tracked as task #119.

Next