HarmonAI

Positioning

HarmonAI is not really a product, and I think our catalog entry flatters it by calling it one. It is — or was — a research lab: a Stability AI-affiliated group, founded around 2022, that released open-source generative audio models for music production. The harmonai.org landing page still says as much ("A Stability AI Lab releasing open-source generative audio tools to make music production more accessible and fun for everyone") and still carries the slogan "AI by musicians, for musicians". But what you actually get when you visit is a static Squarespace page with three value-proposition headings and two links: one to the Harmonai-org GitHub organization, one to a Twitter account. There is no hosted product, no download page, no app, no pricing page — nothing to subscribe to and nothing to log into. Everything of substance lives in three GitHub repositories, and the most recent commit activity across all of them is from April 2024.

That does not make HarmonAI worthless. It made real artifacts that thousands of people used: Dance Diffusion, the training and generation notebooks that were among the first practical ways to train a diffusion model on your own sample library. But a review has to describe what a reader gets today, and today HarmonAI is a dormant lab page pointing at frozen repositories, while the lineage it started — the Stable Audio family — has moved to Stability AI's own GitHub organization and is actively developed there. If you came here looking for an AI music production tool you can use this week, this listing is a signpost, not a stop.

What it actually does

Strip away the landing page and HarmonAI's deliverables are three MIT-licensed repositories, which I enumerated directly through the GitHub API on 24 October 2026:

• sample-generator — "Tools to train a generative model on arbitrary audio samples", 1,116 stars and 168 forks, last pushed 29 April 2024. This is the famous one. It ships two Google Colab notebooks, Dance_Diffusion.ipynb and Finetune_Dance_Diffusion.ipynb, so you can train or fine-tune Dance Diffusion on your own audio without setting up a local ML environment. The README's prerequisites are a line long: Python 3.7+, pip install ..
• oobleck — "open soundstream-ish VAE codecs for downstream neural audio synthesis", 127 stars, last pushed 12 June 2023. Infrastructure for researchers building neural audio codecs.
• audio-diffusion-pytorch-fork — "Audio generation using diffusion models, in PyTorch", 50 stars, last pushed 28 September 2023. A fork of the audio-diffusion PyTorch stack.

All three carry MIT licenses, which is the honest meaning of "free" in our listing: no price, no seat limits, no usage caps, and you may ship what you build with them. The output the lab is best known for is Dance Diffusion itself — a generative audio model family you can train on arbitrary samples (loops, one-shots, full tracks) and use to generate new audio in the style of the training data. The other headline output is Stable Audio, the fast audio-diffusion system HarmonAI unveiled in September 2023: per the Weights & Biases report at the time, it was trained on over 800,000 audio files (music, sound effects, instrumental samples) and could generate 95 seconds of high-quality stereo audio in under one second on an NVIDIA A100. HarmonAI said it planned to open-source versions of the system.

My testing notes

First, the activity check, because it decides how much of this review matters. I queried the GitHub API for the Harmonai-org organization and got exactly three public repositories back, with pushed-at timestamps of 2024-04-29, 2023-09-28 and 2023-06-12. Two and a half years without a push, across the entire org, is not "quiet maintenance" — it is a stopped project. The issues and pull-request tabs are correspondingly thin. Whatever Discord-era community existed around Dance Diffusion in 2022–2023 has moved on; if you open an issue today, there is no evidence anyone is reading.

Second, I walked through sample-generator's README and notebooks as a prospective user. The entry path is two Colab notebooks — one to train from scratch on your samples, one to fine-tune an existing checkpoint. That is a real, working artifact: people made music with these notebooks, and the 1,116-star count reflects genuine use. But the target user is an ML-comfortable producer, not a musician looking for a DAW plugin. There is no UI, no VST, no hosted inference, no curated model zoo. You provision a GPU runtime, point the notebook at a folder of audio, and manage checkpoints yourself. For the ML engineer or the producer who is also technical, that is freedom; for everyone else, it is a wall.

Third, the lineage question. Stable Audio did not die with HarmonAI's activity — it graduated. The Stable Audio 3 repository lives under the Stability-AI GitHub organization (created 19 March 2026, MIT-licensed), with models published on Hugging Face: the stable-audio-3-small-music checkpoint is listed at 433M parameters with a 120-second maximum length and a CPU-capable profile, built on a new Semantic-Acoustic Music Encoder producing 44.1 kHz stereo latents. That is the actively maintained descendant of what HarmonAI started, and for practical purposes it is where the roadmap went. HarmonAI's own page simply never updated to say so.

One more observation worth a line: domain hygiene is poor around this brand. Coverage from 2023 links harmon.ai while the real org page is harmonai.org, and the footer of harmonai.org defers to Stability AI's privacy policy and terms of use — the legal posture of an affiliate lab, not a standalone company. Tracxn lists it as founded in 2022 and describes it as a developer of open-source generative audio tools; that matches what I see.

What it does not do

There is no hosted product at all — no web app, no API, no paid inference tier, nothing you can use without running code. There is no support channel: no Discord link on the page, no email, no issue-triage commitment (the itool-bible listing makes the same call: "free open-source models and code; no hosted product on this site"). There are no updates — the repositories are frozen, and the README's "Todo" list is permanently open. The Colab path depends on Google's runtime environment and third-party package versions from 2023; expect dependency friction before you generate a single sample. Dance Diffusion is sample- and style-bound: it does not write songs, compose to a brief, produce stems on demand, or generate natural speech (the W&B report noted Stable Audio likely had no strong speech generation out of the box). And training anything worthwhile needs a GPU budget and a clean sample library — the models are only as good as the audio you feed them.

Verdict

I respect what HarmonAI did: it open-sourced practical generative audio tooling early, gave it MIT terms, and seeded the Stable Audio line that Stability AI now develops. But reviewing what a reader can pick up today, it is a dormant lab with three frozen repositories, no product surface, no support, and a landing page that has not caught up with reality. For an ML engineer mining permissively licensed audio-diffusion code, the repos remain useful and legitimately free. For anyone shopping for "AI music production unleashed", the honest guidance is to look at Stable Audio 3 under Stability AI, or at maintained tools like MusicGen and the commercial audio generators, instead. I rate HarmonAI 2.3 — the artifacts are real and technically substantive but demand ML skills, which limits the core-functionality score (40%); the price of zero is unbeatable (20%); usability is low with notebooks as the only interface (20%); there is no live support or ecosystem activity (10%); and the reliability signal is a two-and-a-half-year-old last commit with the work continued elsewhere (10%).

Sources

• https://www.harmonai.org/
• https://github.com/Harmonai-org
• https://github.com/Harmonai-org/sample-generator
• https://github.com/Harmonai-org/oobleck
• https://github.com/Harmonai-org/audio-diffusion-pytorch-fork
• https://github.com/Stability-AI/stable-audio-3
• https://wandb.ai/byyoung3/ml-news/reports/Harmonai-s-New-Audio-Diffusion-Model-Stable-Audio--Vmlldzo1Mzg0MDcz
• https://www.audiocipher.com/post/dance-diffusion-harmonai
• https://theaitoolbible.com/tools/harmonai
• https://platform.tracxn.com/a/d/company/59b0ed5be4b017b1013b8b87/harmonai