Introduction

Here's an open music model you can use without wondering whose records it learned from. On October 18, mimetic-a0 landed on Hugging Face and GitHub from Mick Grierson (GitHub micknoise), whose profile lists him as Professor of Computing at the Creative Computing Institute. It's a text-to-audio model for music, trained only on Freesound clips released as CC0 plus music rendered by its own procedural generators. It writes 48 kHz stereo clips up to 16 seconds: drum loops, basslines, chords, melodies, pads, one-shots, and rough full mixes, at the tempo, key, and chords you type. Weights and code are MIT-licensed, and it runs today on an NVIDIA GPU, an Apple Silicon Mac, or a CPU, with a command-line tool and a local web UI.

What Shipped

The model

According to the model card:

  • A 326M-parameter diffusion transformer (rectified flow, 16 layers), 511M parameters in total once you count the audio autoencoder.
  • It works in the latent space of the ACE-Step 1.5 VAE, used frozen, and reads prompts with a frozen T5-base text encoder.
  • Trained from scratch with stock stable-audio-tools on two RTX PRO 6000 cards, 100k steps plus a 20k-step anneal.
  • Two weight files: fp32 at 2.0 GB and fp16 at 1.0 GB.
  • It's labeled version 0.1, an early release, and is also meant as a base for fine-tuning, with updates promised.

The training data

This is the headline. The mix is about 35% Freesound CC0 clips (roughly 276k of them, after a provenance review that threw out 2,762 clips that looked like AI outputs, voice clones, commercial packs, Apple Loops, game, film, or YouTube rips, or named Amen breaks) and 65% synthetic music: 1.53M items rendered by the author's own generators. Each synthetic clip gets a caption computed from the notes actually played, which is why the model follows chords, tempo, and key so literally. Most instruments come from the GeneralUser GS SoundFont, whose author allows use "without restriction" for private or commercial music.

The card is upfront about the soft spots. A CC0 tag is the uploader's own claim, so an unlabeled clip could still slip through, and the SoundFont's author notes the origin of some of its samples is uncertain. That's a lot more transparency than most music models offer.

Speed and control

A 12-second clip at 50 steps takes 0.4 seconds in fp16 on an RTX PRO 6000 and about 6.7 seconds on an M1 Max, and a 3-second clip took 18 seconds on a many-core CPU. On the author's held-out tests it hit the requested tempo 88% of the time and the right pitch on single notes 94% of the time, with chords correct 77% of the time.

Try It

You need uv and git. The first run downloads the weights and the T5 encoder (about 900 MB) for you.

git clone https://github.com/micknoise/mimetic-a0
cd mimetic-a0
uv sync
uv run mimetic-a0-webui

Then open http://127.0.0.1:8799. Besides plain Generate, the web UI has Retake, which regenerates audio you upload, and Repaint, which redoes just a stretch you select on the waveform and keeps the rest. Add --half for the fp16 weights. Owners of RTX 50-series cards need the CUDA 12.8 PyTorch build, and the README has the one-line fix.

From the command line, --bpm and --bars set the length to exact bars:

uv run mimetic-a0 "boom bap drum loop, standard kit, swung, with ghost notes" --bpm 90 --bars 4 -n 4

Prompts follow a set shape: what it is first, then the music, then tempo, key, and length. Here are a few starters in the wording it was trained on:

finger bass bassline, over Cm - Fm - G7, sparse, 95 BPM, C minor, 4 bars
electric piano chords, chords Am7 - D9 - Gmaj7, smooth, sustained, 100 BPM, G major, 2 bars
warm pad held chords, chords Dm7 - G7 - Cmaj7, soft, 80 BPM, C major, 4 bars
marimba arpeggio, chords C - Am - F - G, 120 BPM, C major, 2 bars
808 kit kick drum hit, one-shot, hard

My tip: match the duration to the bars, using seconds = bars × 240 ÷ BPM, and write "90 BPM" rather than "90bpm". The prompting guide lists every instrument, kit, and style it knows, and sticking to that vocabulary makes a real difference.

Why It Matters for Makers

Licensing is the question hanging over every AI music tool, and mimetic-a0 answers it in the most direct way possible: it only trained on audio that was given away or made from scratch. For beatmakers and game audio folks, that makes it a sample generator you can drop into a project with a clear conscience. The trade-offs are honest ones. It's 4/4 only, one, two, or four bars, 16 seconds at most. Instruments sound dry and General MIDI-like, sound effects and foley are weak, and the autoencoder caps the very top end. It also doesn't load in ComfyUI's Stable Audio nodes yet. If you want full songs with vocals, YuE2 or ACE-Step are still your tools. If you want clean, tempo-locked parts you can chop, this is a lovely new option.

Conclusion

mimetic-a0 is a 511M-parameter, MIT-licensed music model trained only on CC0 Freesound clips and synthetic renders, making 48 kHz stereo loops and one-shots up to 16 seconds on a GPU, a Mac, or a CPU. Clone the repo, open the web UI, and ask it for a swung boom bap loop at 90 BPM. Happy looping!

Original source:

https://huggingface.co/micknoisemachine/mimetic-a0

Code and web UI: mimetic-a0 on GitHub · Prompting guide: PROMPTING.md

—Anabel ♡