airadio — local AI radio with MiniMax Music 3

A local AI radio loop: generate ~2-minute songs with MiniMax Music 3 in the background, play them with tail fades, and fill gaps with prerecorded interstitials or library tracks until the next song is ready.

Install as airadio (PyPI-ready; not published yet):

git clone https://github.com/Decentricity/interdimensional-radio.git
cd interdimensional-radio
pip install -e .
airadio

Each new song gets a random two-word title that primes generation and becomes the filename (velvet-garden4289.wav).

Hardware

Tested on NVIDIA GeForce RTX 4070 Ti (12 GB VRAM). This repo uses the INT8-quantized MiniMax Music 3 weights (~11 GB on disk; ~10–11 GB VRAM during a 120s generation).

Profile	GPU VRAM	Model set
Recommended (this repo)	12 GB+	INT8 DiT + pruned INT8 text encoder + DAV VAE
Higher quality	16 GB+	FP16 DiT + pruned BF16/FP16 text encoder + DAV VAE
Minimum	8 GB (tight)	INT8 DiT + pruned INT8 text encoder; shorten --duration

You also need:

• Linux (tested; PipeWire pw-play or ALSA aplay for playback)
• ~15 GB free disk for the three model files
• ffmpeg and ffprobe on PATH
• Python 3.10+
• curl, flock (util-linux)

Quick start

git clone https://github.com/Decentricity/interdimensional-radio.git
cd interdimensional-radio

# 1. Install the package (editable dev install)
pip install -e .

# 2. Install MiniMax Music 3 + ComfyUI (see below)

# 3. Optional: point runtime data at this checkout (auto-detected if library/ exists here)
export AIRADIO_HOME="$PWD"

# 4. Generate interstitial clips (optional but recommended)
python3 bin/generate-interstitials.py music3-both

# 5. Run the radio
airadio

Legacy wrappers bin/interdimensional-radio and bin/music3-warm forward to airadio.

Set AIRADIO_HOME if you keep runtime data (library, interstitials audio, staging) outside the install directory. Runtime state lives under $AIRADIO_HOME/.radio-staging/.

CLI

Command	Purpose
airadio	Start the radio loop
airadio run	Same as above
airadio title	Print random two-word song titles
airadio music3 start	Start warm ComfyUI server
airadio music3 stop	Stop warm server
airadio music3 status	Server + GPU status
airadio music3 gen …	One-off generation

Install MiniMax Music 3

Interdimensional Radio drives generation through ComfyUI with native MiniMax Music 3 nodes. Official docs: ComfyUI MiniMax Music 3 tutorial.

1. ComfyUI (0.33.0+)

ComfyUI must include the MiniMax Music 3 audio nodes. Update to the latest release or nightly if the template is missing.

git clone https://github.com/comfyanonymous/ComfyUI.git ~/.local/share/comfy-music3
cd ~/.local/share/comfy-music3
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Or use the ComfyUI installer and point COMFY_ROOT at that install.

2. Download model weights

Download from the Comfy-Org MiniMax Music 3 Hugging Face repo (Apache 2.0).

Low-VRAM set (used by this repo):

File	Install path under $COMFY_ROOT
minimax_music3_dit_int8_convrot.safetensors (~2.5 GB)	models/diffusion_models/
minimax_music3_text_encoder_pruned_int8_convrot.safetensors (~8.6 GB)	models/text_encoders/
minimax_music3_dav.safetensors (~207 MB)	models/vae/

Example with huggingface-cli:

export COMFY_ROOT="${COMFY_ROOT:-$HOME/.local/share/comfy-music3}"
pip install huggingface_hub
huggingface-cli download Comfy-Org/minimax-music-3_ComfyUI \
  minimax_music3_dit_int8_convrot.safetensors \
  --local-dir "$COMFY_ROOT/models/diffusion_models"
huggingface-cli download Comfy-Org/minimax-music-3_ComfyUI \
  minimax_music3_text_encoder_pruned_int8_convrot.safetensors \
  --local-dir "$COMFY_ROOT/models/text_encoders"
huggingface-cli download Comfy-Org/minimax-music-3_ComfyUI \
  minimax_music3_dav.safetensors \
  --local-dir "$COMFY_ROOT/models/vae"

Restart ComfyUI after placing files. If a loader dropdown is empty, the file is in the wrong folder or still downloading.

3. comfy-cli (generation client)

export MINIMAX_ROOT="${MINIMAX_ROOT:-$HOME/.local/share/minimax-music3}"
mkdir -p "$MINIMAX_ROOT"
python3 -m venv "$MINIMAX_ROOT/venv"
"$MINIMAX_ROOT/venv/bin/pip" install comfy-cli

music3-warm expects the CLI at $MINIMAX_ROOT/venv/bin/comfy. Override with COMFY_BIN if needed.

4. Smoke test

export AIRADIO_HOME=/path/to/interdimensional-radio   # or any data directory
airadio music3 start
airadio music3 gen \
  --lyrics "$(python3 -c 'from airadio import song_title; print(song_title.build_lyrics("Test Song"))')" \
  ...

For a real smoke test, use bundled prompt files under the installed package or write temp caption/lyrics files.

airadio music3 start
# use paths from your AIRADIO_HOME or package data
airadio music3 stop

How it works

┌─────────┐     ┌──────────────────┐     ┌─────────────┐
│ airadio │────▶│ airadio.music3   │────▶│ ComfyUI     │
│ run     │     │ (warm server)    │     │ MiniMax M3  │
└────┬────┘     └──────────────────┘     └─────────────┘
         │
         ├── library/          saved songs (title+seed.wav)
         ├── interstitials/    gap-fill clips
         └── prompts/          caption, lyrics template, title words

1. Startup — pick a random library track as warm-up; start generating the first titled song.
2. Pipeline — while a song plays, the next one generates (~2–3 min for 120s INT8).
3. Fill — if gen is not ready: up to 40s of interstitials, then a full library track; repeat.
4. Archive — each finished song is copied to library/ as {phrase}{seed}.wav with catalog metadata.

Scripts

Entry point	Purpose
airadio	Main radio loop + subcommands
bin/generate-interstitials.py	Batch-generate station IDs, ads, DJ chatter

Bundled in the package: prompts, title word lists, ComfyUI workflow builder, interstitial scripts/prompts.

Environment variables

Variable	Default	Meaning
AIRADIO_HOME	~/.local/share/airadio (or cwd if library/ exists)	Runtime data root
IDR_ROOT	(alias for AIRADIO_HOME)	Legacy name
COMFY_ROOT	~/.local/share/comfy-music3	ComfyUI install
MINIMAX_ROOT	~/.local/share/minimax-music3	comfy-cli venv
COMFY_BIN	$MINIMAX_ROOT/venv/bin/comfy	comfy-cli binary
AIRADIO_PROMPTS	bundled package prompts	Override prompt directory

Prompts and titles

• prompts/normie-control.caption.txt — production metadata (genre-neutral pop).
• prompts/normie-control.lyrics.template.txt — lyrics scaffold; {word1} and {word2} from the title woven into verse lines (title stays in caption only).
• prompts/song-title-words.json — 32×32 word lists for grammatical two-word titles.

Edit the word lists or caption to steer the station sound. The radio writes per-generation caption/lyrics files under .radio-staging/prompts/.

Interstitials

See interstitials/README.md. Generated WAVs are gitignored; scripts and prompts are included. Regenerate with:

python3 bin/generate-interstitials.py music3-both

License

Application code in this repository: MIT (see LICENSE).

MiniMax Music 3 model weights are distributed separately under Apache 2.0 by MiniMax / Comfy-Org. Download and use of the weights is subject to their license terms.