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: Together AI
providerId: together-ai
created: '2026-05-08'
# Provenance stamped 2026-08-11: this artifact was written by the API Evangelist
# bulk sweep dated 2026-05-08, not harvested from the provider. See roadmap#35.
method: generated
modified: '2026-05-08'
reconciled: false
tags:
- AI
- LLM
- Inference
- Open Source
- Fine-tuning
- Rate Limiting
- Quotas
- Throttling
description: >-
  Together AI enforces per-account rate limits on serverless inference that
  vary by model and account tier (Build / Scale / Enterprise as account
  spend/credit grows). Limits include requests-per-minute (RPM) and
  tokens-per-minute (TPM) per model. Specific per-model values are not
  reconciled in this artifact - see the Together console for active limits
  on your account.
notes: >-
  Per-model RPM/TPM limits are surfaced in the rate-limits page of the user
  console. Dedicated and reserved deployments bypass the shared serverless
  limits within the provisioned capacity.
sources:
- https://docs.together.ai/docs/rate-limits
- https://www.together.ai/pricing
responseCodes:
  throttled: 429
limits:
- name: Requests Per Minute (RPM)
  scope: account
  metric: requests
  limit: see provider documentation
  notes: Per-model RPM, varies by tier and model. Pending reconciliation.
- name: Tokens Per Minute (TPM)
  scope: account
  metric: tokens
  limit: see provider documentation
  notes: Per-model TPM, varies by tier and model. Pending reconciliation.
- name: Concurrent Fine-Tuning Jobs
  scope: account
  metric: jobs
  limit: see provider documentation
  notes: Concurrency cap on parallel fine-tuning jobs.
- name: Batch Job Size / Concurrency
  scope: account
  metric: jobs
  limit: see provider documentation
  notes: Batch jobs are queued and do not consume serverless RPM/TPM directly.
- name: Dedicated Endpoints
  scope: endpoint
  metric: requests
  limit: bounded by provisioned GPU capacity
  notes: Throughput is determined by the dedicated hardware sizing.
policies:
- name: Tiered Limits
  description: Limits scale up automatically with account spend / credit balance and via Enterprise agreements.
- name: Backoff Strategy
  description: Clients should implement exponential backoff with jitter and honor any Retry-After header.
maintainers:
- FN: Kin Lane
  email: kin@apievangelist.com