apiCommonsRateLimits: '0.1'
provider:
  name: NVIDIA NIM
  id: nvidia-nim
  url: https://docs.api.nvidia.com/nim/reference/limits
  sources:
  - https://docs.api.nvidia.com/nim/reference/limits
  - https://build.nvidia.com
  - https://docs.nvidia.com/nim/large-language-models/latest/configuration.html
scope:
  description: Limits below apply to the NVIDIA-hosted NIM endpoint at integrate.api.nvidia.com. Self-hosted NIM
    containers are governed by the GPU capacity of the host and a configurable concurrency limit per container.
policies:
- id: hosted-developer-rpm
  name: Hosted Developer RPM
  scope: per-api-key
  surface: https://integrate.api.nvidia.com
  metric: requests
  limit: 40
  window: 1m
  description: Free developer tier soft rate limit on /v1/chat/completions, /v1/completions, /v1/embeddings, and /v1/ranking.
- id: hosted-developer-credits
  name: Hosted Developer Credits
  scope: per-account
  surface: https://integrate.api.nvidia.com
  metric: requests
  limit: 1000
  window: signup
  description: 1,000 free inference credits granted on NVIDIA Developer Program signup, consumed across all hosted models.
- id: hosted-concurrent
  name: Hosted Concurrent Requests
  scope: per-api-key
  surface: https://integrate.api.nvidia.com
  metric: concurrent-requests
  limit: 5
  window: instant
  description: Concurrent in-flight requests against the shared hosted endpoint.
- id: hosted-streaming-keepalive
  name: SSE Keepalive
  scope: per-stream
  surface: https://integrate.api.nvidia.com
  metric: idle-time
  limit: 60
  window: 60s
  description: Streaming connections idle longer than ~60s may be closed by the gateway.
- id: hosted-max-tokens
  name: Per-request Output Cap
  scope: per-request
  surface: https://integrate.api.nvidia.com
  metric: max_tokens
  limit: 4096
  window: per-request
  description: Per-request output token cap on the hosted endpoint (model-dependent; some models support higher).
- id: hosted-input-size
  name: Hosted Input Size
  scope: per-request
  surface: https://integrate.api.nvidia.com
  metric: input_tokens
  limit: 128000
  window: per-request
  description: Maximum input tokens on the hosted endpoint (model-dependent — long-context Llama 3.1/3.3 and Nemotron variants accept up to ~128K).
- id: self-hosted-concurrent
  name: Self-hosted Container Concurrency
  scope: per-container
  surface: http://localhost:8000
  metric: concurrent-requests
  limit: unbounded
  window: instant
  description: Self-hosted NIM containers accept as many concurrent requests as the GPU and TensorRT-LLM batching configuration allow. Operators tune via NIM_MAX_BATCH_SIZE and similar env vars.
modified: '2026-05-25'