audio.cpp with minimax-music-ui

[MerijnHendriks]

Hey everyone,

Sorry if this is a dumb question, very new to using audio.cpp!

I'm trying to make audio.cpp with minimax work together with minimax-music-ui:
https://github.com/adambenhassen/minimax-music-ui

When I used the model from the model downloader:

python model_manager_v2.py install minimax_music3_q4_0

And use the following config:

{
  "host": "127.0.0.1",
  "port": 7862,
  "backend": "cpu",
  "device": 0,
  "threads": 6,
  "lazy_load": true,
  "models": [
    {
      "id": "minimax-music3",
      "family": "minimax_music3",
      "path": "C:/LLMs/inference/models/MiniMax-Music3-GGUF",
      "task": "gen",
      "mode": "offline"
    }
  ]
}

I get this error when generating using minimax-music-ui:

upstream 500 on /v1/audio/speech: {"error":{"message":"builtin model spec for family 'minimax_music3'.packages[0].precision: unknown precision 'q4_0'","type":"server_error"}}

Using the build-in UI from audio.cpp does work however.

Any ideas what I might be doing wrong?

[adambenhassen]

Hi! minimax-music-ui maintainer here.

Before you dive into specifics, just want to let you know that you have done nothing wrong. It's a bug and I'm sure the next audio.cpp release will fix it.

Now to the specifics. This error isn't coming from the UI. It's raised by audio.cpp's spec validator (engine::model_spec::validate_spec), and it points to a version mismatch on your install: your model_manager_v2.py is new enough to install minimax_music3_q4_0, but your audiocpp_server binary predates q4_0 support.

Why the built-in UI works but this UI doesn't: audio.cpp validates the model spec lazily on the model's first request through the OpenAI-compatible endpoint, and that path resolves through the compiled-in builtin spec table. The bundled web UI loads via the on-disk model_specs/*.json path instead, so it never hits that check.

You can confirm the UI isn't a factor by reproducing with plain curl:

curl <server>/v1/models # note the advertised id, the UI sends this back verbatim

then:

curl -X POST <server>/v1/audio/speech -H 'Content-Type: application/json' -d '{"model":"<that id>","input":"test","instructions":"pop","response_format":"wav","max_new_tokens":250,"stream":false}'

Same error: it's entirely server-side.

Two ways to fix it:

1. Patch/rebuild audio.cpp so the server binary knows q4_0 (recent commits include it).
2. Or reinstall the model at a precision your build accepts
3. Wait until it's fixed by the devs and download the new release binary

For the UI side there's nothing to change, it sends the standard {model, input, instructions, response_format, max_new_tokens, stream} payload with the model id your server itself advertises on /v1/models.

[0xShug0]

@MerijnHendriks @adambenhassen audio.cpp currently routes Music3 via /v1/tasks/run, not /v1/audio/speech. The native audio.cpp UI uses /v1/tasks/run for Music3, which is why it works there.

All (almost) models in audio/music generation, audio tools, speech analysis, and voice conversion route only through /v1/tasks/run

I’ll still relax the schema validation for the speech endpoint, but the correct endpoint for Music3 integrations is /v1/tasks/run.