Serverless Inference

Meshive Serverless turns a model into a private, OpenAI-compatible endpoint in a few clicks. You never pick a GPU — set a price cap and a replica range, and the cheapest GPU that fits is assigned (and re-assigned) automatically.

Model (official catalog or your Hugging Face repo)
        │  Deploy — price cap + min/max replicas
        ▼
Serving deployment ──▶ https://<your-deployment>.inference.meshive.ai/v1
        │                      OpenAI SDK–compatible
        ▼
Billed per GPU-hour of running replicas

What you can run

Modality	Endpoints	Mode
Chat / text	/v1/chat/completions, /v1/completions	Sync, streaming supported
Embeddings	/v1/embeddings	Sync
Image	/v1/images/generations, /v1/images/edits	Sync (30s budget) or async
Video	/v1/videos, /v1/videos/generations	Always async

One deployment = one model = one endpoint URL. If it works in the Playground, the same request works through the API.

How you’re billed

• GPU-hour × running replicas. Nothing else — your own deployments carry no per-request or per-token charges. (Only shared models, which you don’t deploy, bill per token.)
• Every GPU candidate’s hourly price is shown before you confirm a deployment.
• Your price cap is a hard ceiling per replica-hour — never exceeded.
• The Serving page shows each deployment’s live $/hr, covering exactly the replicas currently billed (when billing starts).

The seven consoles

Everything lives under Serverless in your workspace console:

Page	What it’s for
Models	Official catalog, registering your own HF models, Playground, API keys
Serving	Deploying and operating endpoints — replicas, autoscaling, price cap, metrics
Shared	Models other teams host — call them per token with no deployment; or open your own idle capacity and earn
Tasks	Running a Python script once, to completion — batch jobs, fine-tuning, preprocessing
Requests	Tracking every async request — status, exact failure cause, results
Artifacts	Generated images & videos — gallery, storage, retention
Webhooks	Push notifications when async requests finish

Sync and async

Path	Covers	Notes
Sync	Chat, completions, embeddings, images (default)	Images have a 30-second budget → 504 past it
Async	Images with ?async=true, all videos	Accepted instantly with a request id → poll, webhook, or console

Note

Async is also where failures get precise: a failed async request carries an exact error_code (e.g. vram_exceeded) on the Requests page. Sync 5xx errors can’t always be classified.

Authentication

Every endpoint authenticates with a workspace API key, created on the Models page → API Keys tab (details). The plaintext is shown once — only a hash is stored.

Authorization: Bearer mk-...