MiniMax-Music3 Just Dropped: The Open-Source Music Gen Model That Actually Slaps

MiniMax dropped Music3 yesterday and the subreddit is losing its collective mind. I get it. This thing does 48kHz stereo generation with a 1M token context window. That’s not a typo. For comparison, MusicGen was stuck at 32kHz mono and felt like it was generating through a tin can.

The thread’s top comment nailed it: “Finally a music model that doesn’t sound like a MIDI file having a seizure.” Brutal, accurate, and honestly the best endorsement I’ve seen for an audio model in months.

What Actually Changed #

Music3 isn’t just an incremental bump. The architecture shift to a latent diffusion transformer means you’re getting coherent structure over long horizons. The demo clips show actual verse-chorus-verse progression, not just 30 seconds of vibes that collapse into noise.

The 1M context is the real headline though. You can feed it an entire song and ask for a remix, or give it a 10-minute podcast and get a theme track that matches the emotional arc. That’s genuinely new territory for open weights.

The Setup: Not For The Faint Of Heart #

Here’s where the rubber meets the road. This thing is hungry. The community consensus from the thread is that you need at least 24GB VRAM for the 7B model at full precision. The 3B variant runs on 12GB but you’re sacrificing quality noticeably.

git clone https://github.com/MiniMax-AI/Music3
cd Music3
pip install -r requirements.txt

That’s the easy part. The real pain point is the weights. They’re on Hugging Face but the download is 14GB for the 7B. If you’re on a Hetzner box with their 1Gbps uplink, that’s a coffee break. On DigitalOcean’s standard bandwidth, you’re looking at 20 minutes of staring at a progress bar.

Generation: The Actual Commands #

Once you’re set up, generation is refreshingly simple:

from music3 import MusicGenerator

gen = MusicGenerator(model_size="7b", device="cuda")
prompt = "melancholic synthwave, 110 BPM, with a driving bassline and ethereal pads"
audio = gen.generate(prompt, duration=30, stereo=True)
audio.save("output.wav")

That’s it. No prompt engineering gymnastics, no negative prompts required. The model handles musical intent surprisingly well. One user in the thread generated a “lofi hip hop beat with rain sounds” and the output had actual vinyl crackle baked in. That’s the kind of detail that makes this worth the VRAM.

Where It Falls Apart #

I love this tool but it has one fatal flaw: the licensing is murky. The weights are open but the training data provenance is vague. If you’re planning to release commercial music, you’re in a gray zone that could bite you later. The thread has a whole comment chain debating this and nobody has a definitive answer.

Also, the CPU inference path is basically a joke. Someone tried running the 3B on a MacBook M2 and reported 45 seconds per second of audio. That’s not usable for iteration. You need a real GPU or you’re going to hate life.

The Verdict (Such As It Is) #

If you have the hardware, Music3 is the best open-source music generation model available right now. Period. The quality gap between this and the previous generation is like comparing Stable Diffusion 1.5 to SDXL. It’s not close.

If you’re on consumer hardware, wait for the quantized versions. The community is already working on GGUF conversions and someone in the thread claims a 4-bit version should hit 8GB VRAM. Your mileage may vary, but I’d bet on that landing within the week.

FAQ #

Q: Can I use Music3 for commercial projects? A: The weights are open but the training data license is unclear. Check the official repo for the latest terms before shipping anything commercial.

Q: What’s the minimum hardware requirement? A: 12GB VRAM for the 3B model, 24GB for the 7B. CPU inference is technically possible but painfully slow — expect 45+ seconds per second of audio.

Q: Does it support text-to-music and music-to-music? A: Yes. The 1M context window lets you feed existing audio for remixing, extension, or style transfer. That’s the killer feature.