Rate limits

Updated 2026-10-29.

Every limit returns an error code. Retryable errors include a Retry-After header (seconds).

Limits

Limit	Value	Error	Reset
Calls per key	200 per UTC day (default)	429 `daily_limit`	00:00 UTC
Tokens per member	300,000 per UTC day	429 `daily_tokens`	00:00 UTC
Open calls per member	64	429 `too_many_streams`	call end
Open calls, all keys	96	503 `device_busy`	call end
Request body	32 MB	413 `too_large`	none
Non-streamed response	about 100 s	524	none
Tokens per request	`context_length`	400	none
Model load wait	15 min	503 `model_loading`	none
GPU queue wait	900 s	503 `gpu-busy`	none

Member limits apply to the session routes. context_length is per model, in GET /v1/models. Open calls count toward the daily call limit. GET /v1/models does not count. Keys have no token limit.

Usage

Citadel does not return the remaining call count. A 429 daily_limit response includes limit and used.

Recommendations

Set max_tokens on every request.
Use the loaded model. A model switch takes 2 to 5 minutes.
Limit concurrency to 64 on qwen3.8-27b. Requests above 64 wait in the engine.
For a higher daily limit, ask the owner of albastov.com.