Repository navigation
feat: OpenAI Live model support (Bi-directional voice models) #33175
Description
Before Submitting
- I searched open and closed issues and discussions for an existing request.
- I checked whether this already exists on the
devbranch or latest source. - I understand that maintainers want a well-written issue or discussion before any code pull request.
- This request is not a security vulnerability.
Problem
I love the realtime models that were added. This is huge!
Though time has passed and now it seems openai focusses on their new branch of OpenAI live models.
It would be very cool if they could also be added as they have superior interaction quality.
Desired Behavior
Support bi-directional voice models.
Why This Matters
I suspect that live models will be the future, having to pause for an LLM to reply will perhaps be a relict of the future in voice conversations. It doesnt give that much of a benefit other than being in a more natural conversation, but I also expect the model development to pivot more into these types of model architectures, resulting in that realtime models will be a bit left behind soon.
Also note that there are other models that support this. Thereby enabling the models to map to an underlying generic model instead of adopting openai's api may be beneficial. E.g. Qwen3-Omni, Moshi, GLM-4-Voice
Examples or References
No response
Alternatives or Workarounds
No response
Activity
owui-terminator commentedon Oct 25, 2026
🔍 Related Issues Found
I found some existing issues that might be related. Please check if any of these are duplicates or contain helpful solutions:
🟣 #5894 feat: openai real-time api
This is the earlier request for OpenAI real-time API support in Open WebUI, which is the same feature area as the new issue asking for OpenAI Live / bi-directional voice model support.
by stevenbaert ·
non-core🟣 #32343 feat: Azure OpenAI Realtime support for Voice Call mode
This issue asks for Azure OpenAI Realtime support in Voice Call mode, closely overlapping with the new request’s broader goal of supporting next-generation OpenAI-hosted live/realtime voice models.
by msi-martinsen ·
enhancement🟢 #23357 feat: send audio and video directly to natively multimodal models instead of forcing STT transcription
This is about sending audio directly to natively multimodal models instead of forcing STT transcription, which is closely related to the new issue’s emphasis on bi-directional voice models and native audio interaction.
by TomTheWise ·
enhancement💡 If your issue is a duplicate, please close it and add any additional details to the existing issue instead.
This comment was generated automatically. React with 👍 if helpful, 👎 if not.
DavidProf commentedon Oct 31, 2026
I tried using realtime with LFM3-Omni-2B and llama.cpp build 12012, on open web ui v0.12.5, but I always get the error
Not sure if I'm doing something wrong and thought this issue was related...
Config for realtime:
Voice Model
LiquidAI/LFM3-Omni-2B-GGUF
Input Transcription Model
LiquidAI/LFM3-Omni-2B-GGUF
Same as the id comming from /models on llama
Pl8tinium commentedon Oct 31, 2026
@DavidProf this is an omni model, not a realtime model. These full duplex models can listen and push audio at the same time.
This is what this issue requests support for.
Realtime models on the other hand continously listen while talking and can thereby, on a pause, immediately reply to the user. They wont interrupt the user, affirm, or do anything else while the user is talking though.
I think the development will go to live models in the future.
DavidProf commentedon Oct 31, 2026
@Pl8tinium Am I understanding something wrong? Doesn't this model support realtime?
As said here? https://huggingface.co/LiquidAI/LFM3-Omni-2B#inference
Pl8tinium commentedon Oct 31, 2026
yea as i said, realtime is different than omni/ full duplex models. openwebui supports realtime thus far.