specification: API Commons Rate Limits
specificationVersion: '0.1'
schema: https://raw.githubusercontent.com/api-evangelist/interface-research/main/schema/api-commons.yml#/$defs/RateLimits
provider: Cerebrium
providerId: cerebrium
created: '2026-06-20'
modified: '2026-06-20'
reconciled: false
tags:
- AI
- GPU
- Serverless
- Inference
- ML Infrastructure
- Rate Limiting
- Quotas
- Throttling
description: >-
  Cerebrium does not throttle deployed function endpoints with classic
  per-minute request quotas; instead, throughput is governed by autoscaling
  GPU/CPU concurrency limits per app and per account plan tier. The number of
  concurrent replicas an app can scale to is configured per deployment
  (min/max instances) and capped by the plan tier, with Enterprise offering
  unlimited GPU concurrency. Async runs are bounded by a maximum execution
  window (up to 12 hours) and a configurable response grace period. Specific
  per-account concurrency caps are not reconciled in this artifact.
notes: >-
  Verify per-app min/max replica settings, plan-tier concurrency caps, and async
  execution limits in the Cerebrium dashboard and docs on reconciliation;
  values change as accounts move between Hobby, Standard, and Enterprise tiers.
sources:
- https://www.cerebrium.ai/docs
- https://www.cerebrium.ai/docs/cerebrium/endpoints/async
- https://www.cerebrium.ai/pricing
responseCodes:
  throttled: 429
limits:
- name: Concurrent Replicas (per app)
  scope: app
  metric: instances
  limit: see provider documentation
  notes: Configured via min/max instances in cerebrium.toml; capped by plan tier.
- name: GPU Concurrency (per account)
  scope: account
  metric: gpu_instances
  limit: see provider documentation
  notes: Bounded by plan tier; Enterprise offers unlimited GPU concurrency.
- name: Async Execution Window
  scope: request
  metric: seconds
  limit: up to 12 hours
  notes: Async runs are bounded by response_grace_period (default 15 minutes, up to 12 hours).
- name: Cold Start / Scale-to-Zero
  scope: app
  metric: instances
  limit: scales to zero when idle
  notes: Apps scale down to zero idle instances; cold-start latency applies on first request after idle.
policies:
- name: Tiered Concurrency
  description: Concurrency caps raise as accounts move from Hobby to Standard to Enterprise (unlimited GPU concurrency).
- name: Autoscaling
  description: Apps autoscale replicas between configured min and max based on incoming traffic.
- name: Backoff Strategy
  description: Clients should implement exponential backoff with jitter and honor Retry-After on any throttling responses.
maintainers:
- FN: Kin Lane
  email: kin@apievangelist.com