Deepgram Self-Hosted 261110: unsupported language hints now fail with HTTP 400 (November 10, 2026)
Self-hosted release 261110 rejects language_hint on models that don't support it, instead of ignoring it. Remove those hints first.
What changed
Deepgram Self-Hosted release 261110 came out on November 10, 2026. Deepgram marks one change as action required: a language_hint sent to a multilingual model that does not support hints, such as Nova-3 Multilingual, now returns HTTP 400 INVALID_QUERY_PARAMETER. Before, the hint was ignored without a word.
Flux speech-to-text on /v2/listen also checks hints now. An unsupported language_hint fails with HTTP 400 when the stream opens, not after the first audio. A Configure message with an unsupported hint returns ConfigureFailure and the stream keeps its old settings.
Other changes in the same release:
- The
dg-tokenheader is no longer a supported way to verify callbacks. This release still sends it, but you should move to Basic Auth by putting credentials in the callback URL. For speech-to-text you can also pass a secret in Extra Metadata and reject callbacks wheremetadata.extra.callback_secretis missing or wrong. - Models loaded at startup through
[preload_models]now use half precision on supported GPUs, which uses less GPU memory. - The Engine image needs NVIDIA driver 580 or later with the open kernel module.
Prerequisites
- A Deepgram self-hosted deployment you are about to move to release 261110
- Access to the code or config that builds your
/v1/listenand/v2/listenrequests
Steps
- Search your code for
language_hint. For each request that uses a multilingual model or Flux, check that the model supports hints. Remove the hint where it does not. - If your callback handler checks the
dg-tokenheader, switch it to Basic Auth credentials in the callback URL, or to a secret inmetadata.extra.callback_secret. - Check the GPU host runs NVIDIA driver 580 or later with the open kernel module.
- Move the images to the
release-261110tags:self-hosted-api,self-hosted-engine,self-hosted-license-proxyandself-hosted-billing.
Test it
Send one request to a multilingual model with a language_hint it does not support. You should get HTTP 400 INVALID_QUERY_PARAMETER. Remove the hint and send it again: it should return a transcript. Then trigger one callback and confirm your handler accepts it without .
