Hugging Face's logo

MiniMaxAI
/
MiniMax-Music3

Text-to-Audio
Diffusers
Safetensors
PyTorch
minimax_music3
music-generation
text-to-music
sglang-omni
Model card Files
xet
 Community
25

Has anyone gotten a consistently vocal-free track out of the open weights?

#23
by fe2-o3 - opened
Discussion
fe2-o3

We run MiniMax-Music3 on a native Metal/MLX engine (mlx-serve) and it sounds great. We cannot get it to stop singing.

The hosted API has is_instrumental. The open weights have no such parameter — scripts/end_to_end/minimax_ttm_test.py posts only input and instructions — so whatever it does has to be expressible in those two text fields.

What we tried. Same seed and instructions each time, on the released weights at 8-bit:

input (lyrics) instructions result
[Instrumental] prompt wordless vocalizations
[Instrumental] prompt + "Instrumental only: no vocals, no singing, no lyrics." wordless vocalizations
[Inst] prompt + same clause wordless vocalizations
empty ([start] alone) prompt + same clause wordless vocalizations

The instructions already opened with "Instrumental ambient field-recording piece, freely paced, no vocals." — so the request said "no vocals" three ways at once, in the genre, in an explicit clause, and via the lyric tag.

No intelligible lyrics in any take, but clear sung vocal texture throughout, and one opened with an audible word. The empty-lyrics row is the interesting one: with no lyric text at all, it still sings.

One partial result worth mentioning: a take at BPM 167 came out genuinely vocal-free, where the same prompt at 60 BPM did not. Single sample, seed not controlled, so we are not claiming it.

Ruled out already, so nobody has to suggest it:

  • No instrumental special token — the 32 added tokens are <|caption_start/end|>, <|lyrics_start/end|>, <|audio_start/end|>, <|audio_cfg|> and stock Qwen3 markers.
  • tokenizer/chat_template.jinja in the release is the stock Qwen3 tool-calling template, no music path.
  • Our prompt assembly matches the reference, and the engine matches fp32 at prefill cos 0.9999 / DiT velocity 0.999 / vocoder 1.000000, so this is not a numerics bug on our side.

Two questions:

  1. Has anyone gotten consistently instrumental output — no vocalizations at all — from these weights? If so, what did your input and instructions look like?
  2. Does is_instrumental map to something in those two fields at all, or does it select a different conditioning path / checkpoint that is not in this release? If it is text, what does the serving stack put there?

Happy to run any prompt shape you want and post the audio.
Alborz

I saw a similar problem when trying to create an instrumental break between two vocal sections.

I tested the released MiniMax Music3 model with SGLang-Omni using the official prompt format. I ran nine matching seeds for three approaches:

  • Asking for an eight-second instrumental break: 0/9 worked
  • Asking for three complete instrumental bars: 0/9 worked
  • Asking for three bars plus a no-vocals tag: 1/9 worked

I counted a result as successful when it had a 6–10-second vocal-free gap between the verses and was at least 80 seconds long.

Increasing max_new_tokens from 2100 to 2400 did not help. Seven of the nine outputs were exactly the same. The other two only became longer.

I have not tested the maximum 9000-frame setting yet, so the shorter limit may still be part of the problem.

Could MiniMax clarify:

  1. Does the hosted is_instrumental option use something missing from the open weights?
  2. What is the recommended way to request an instrumental break between vocal sections?
  3. Should I use the full 9,000-frame limit even when I want a song around 84 seconds?

I can share my prompts, measurements, and audio samples if helpful.

Sign up or log in to comment