episod/minimax-music3-fiona-crapple-blackhole

minimax-music3-fiona-crapple on Tenstorrent p150 (companion)

Derived from MiniMaxAI/MiniMax-Music3 by MiniMax, with the language-model adapter terminusresearch/minimax-music3-lm-lora-fiona-crapple (regularised/checkpoint-750) merged into the weights at strength 1.0. See the original model cards for training and evaluation details.

MiniMax Music 3, a lyrics-and-caption conditioned music generation model (8B Qwen3-based language model, 0.6B depth decoder for the residual RVQ codebooks, 2.4B flow-matching DiT and a vocoder), with a rank-64 adapter merged into the language model. The adapter was trained on nine tracks (about 52 minutes) in one style and answers to the trigger phrase fiona crapple. The server makes one song at a time and returns 32 kHz stereo audio through the model's own HTTP job API.

Runs on Tenstorrent p150 (mesh P150).

Companion repository for the minimax-music3-blackhole container package: a launcher that serves it with the merged weights, and a Gradio demo.

At a glance

Hardware	p150
License	minimax-music3-community-license

Intended use

Generating songs from lyrics and a text description of the music. Putting fiona crapple in the description steers the output toward the adapter's training style, which its author describes as an intimate cafe singer-songwriter sound with jazzy piano, upright bass, brushed drums and a raw female vocal.

Prerequisites

• The Tenstorrent CLI, tt: uv tool install tenstorrent (or use tt-model alone, see the Quickstart)
• Docker
• Tenstorrent hardware: p150
• About 27 GB of free disk for the merged weights and 4 GB for the container image
• Python 3.12 for the demo (pip install -r requirements.txt)

Quickstart

This repository is a companion to the catalog package. It holds a launcher and a Gradio demo, and the weights are in episod/minimax-music3-fiona-crapple.

git clone https://huggingface.co/episod/minimax-music3-fiona-crapple-blackhole
cd minimax-music3-fiona-crapple-blackhole
uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -r requirements.txt
gozer run --chips 1 -- .venv/bin/python serve/serve_music.py --ui     # without gozer: drop `gozer run --chips 1 --`

The launcher installs the catalog image (tt-model pull jashansinghTT/minimax-music3-blackhole) if it is missing, downloads the weights, starts the server on port 20000 and the Gradio page on port 7871.

Capabilities

This package exposes the model's own HTTP server on port 20000, not an OpenAI-compatible chat API — its request and response shapes are the model's own. See the author's notes for the payload it expects.

curl -s localhost:20000/v1/music/jobs -H 'Content-Type: application/json' -d '{
  "input": "[Verse 1]\nThe kettle sings at half past three\n[Chorus]\nStay a little longer",
  "instructions": "fiona crapple, jazzy piano, raw female vocal, upright bass, brushed drums, intimate",
  "audio_duration": 20, "seed": 11, "response_format": "flac"}'
# -> {"id": "...", "status": "queued", "queue_position": 0}
curl -s localhost:20000/v1/music/jobs/<id>            # status, frames_done, stats when done
curl -s localhost:20000/v1/music/jobs/<id>/audio -o song.flac

audio_duration is in seconds (25 audio frames per second). The same seed, lyrics and description gave the same file: a repeated request on one server, and the same request on a server on the other board, each returned a bit-identical file (one pair each). In a test, 9 jobs submitted at once were accepted and a tenth got HTTP 429.

Expected performance

eval	this package	reference	prompts
Quality evaluation against MiniMax-Music3	not run	not run	0

profile	audio length	concurrency	N	compute time (s)	per second of audio (s)	language model (s)	depth decoder (s)	diffusion (s)	vocoder (s)
p150	20 s	1	20	65.6	3.3	18.2	19.2	24.1	16.0

Medians of the server's own job timings for 20 requests, one song at a time, on one chip of a p300c board (QuietBox 2). The compute time ranged from 64.8 to 68.7 s. One request finished at 16.6 s because the server reported the song as ended. The server was healthy 90 to 172 s after its container started (4 starts).

Limitations

• One song is generated at a time. In a test, 9 jobs were accepted at once (one running, eight waiting) and a tenth got HTTP 429.
• Slower than real time: about 3.3 s of compute per second of audio on one chip.
• Only the p150 profile on one chip is tested, and only with 20 s requests. The server allows up to 9000 frames (6 minutes).
• A song ends when its lyrics are finished, which can be before the requested length: a two-sentence lyric made 5.2 s of a requested 20 s, and 1 of 20 test requests with full lyrics ended at 16.6 s of 20 s. The page says so when it happens.
• No quality evaluation against the base model has been run. The adapter was trained on nine tracks in one style; its effect without the trigger phrase is not measured.
• Not for imitating a specific real artist's voice or presenting generated audio as a recording of a real person.
• On a host where loading a cached tensor of about 32 MiB or more never returns (seen with Linux 7.0.0-38, KMD 2.11.0), the launcher runs the container with its tensor cache disabled.
• The server has no authentication; run it only on a network you trust.

Licensing

Model weights: MiniMax-Music3 Community License; products built with the weights must display "MiniMax-Music3". The merged adapter is Apache-2.0. Launcher and demo code: Apache-2.0.

Related packages

• jashansinghTT/minimax-music3-blackhole: the unmodified MiniMax-Music3 container package this repository's launcher runs.
• episod/minimax-music3-fiona-crapple: the merged weights.

Feedback

Questions or problems with this package: open a discussion at https://huggingface.co/episod/minimax-music3-fiona-crapple-blackhole/discussions — that is what reaches its author. A problem with the tt tooling itself: tt report issue (collects your environment and opens a prefilled issue against tenstorrent/tt-cli). Product feedback: [email protected].

Provenance

What this repository runs:

component	source
container image	sha256:2da7f443622b939c3321fc092b7a0e6948911d78d72cfac53f2749564960451f, from jashansinghTT/minimax-music3-blackhole (model code digest 9a9c37114f86ca58, sha256, first 16 hex digits)
base weights	MiniMaxAI/MiniMax-Music3 at fbdf52fbaaca799592917417eb05f1899f1255ec
adapter	terminusresearch/minimax-music3-lm-lora-fiona-crapple, regularised/checkpoint-750, rank 64, alpha 64
merge	W' = W + strength * (alpha / rank) * B @ A on the 252 language-model matrices the adapter targets, strength 1.0, stored in bf16; every other weight file is the base's (serve/merge_lora.py)