Skip to content
TroubleshootingKimi · Troubleshooting

Kimi · Troubleshooting

Check the model ID, reserve, streaming, and temporary limits.

Quick diagnosis

  1. Check GET /v1/models with the same gd-… key.
  2. Copy one of the four Kimi model IDs without changes.
  3. Make a minimal non-stream request to POST /v1/chat/completions with a small max_tokens.
  4. After a regular response, test stream: true.
  5. For 429, read the exact error.code: causes and retry timing differ.

Common cases

SymptomCheck
404 model_not_foundexact model ID with no aliases or spaces
402 request_reserve_exceeds_creditthe reserve also covers the request input; check the request size and max_tokens — without it the output part uses 32,000
Usage appears twice in a streamthis is native choice usage plus the separate frame with include_usage; do not sum them
429 rate_limit_exceededwait 60–90 seconds; this is not exhausted plan quota
429 kimi_queue_fullretry with a delay
503 kimi_unavailableretry later
404 file_not_found for a fresh filea file lives 24 hours and is visible only to the key that uploaded it
415 unsupported_file_type on uploadeight formats are supported; images, audio, and video are not accepted
400 invalid_managed_service_requestthe search and fetch body must carry exactly one field — query or url — and the body itself must be valid JSON
422 managed_service_request_failedthe site itself, or the search itself, refused. Repeating is pointless and there is no Retry-After: use another address
402 insufficient_credit on DELETE /v1/files/{file_id}files are free, but zero remaining credit closes deletion too; top the key up, or wait for expiry after 24 hours
A second input_file with the same file_id vanishedthe substitution happens once per request; to point at the same document again, say so in words in your own text
The 50 MiB quota ran out on a couple of filesspace is counted on the encrypted record together with the extracted text, which for plain text is about 3.6 times the file size

When to contact support

Send the [code · req_…] tail from the error text (or the request_id), the endpoint, model id, stream true/false and the exact time with its zone. Do not send the full key, the prompt or the answer.

Documentation sections

On this page