HeartMuLa
HeartMuLa/heartlibReviewed by AI
Checked by RepoPicks Radar on
What is HeartMuLa?
HeartMuLa is a family of open music models from MuLa Labs. The released 3B model writes a song with vocals from a lyrics file and a list of tags such as "piano,happy,wedding"; other pieces transcribe lyrics from audio and rebuild 48 kHz stereo sound. You can try the team's hosted demo page in a browser. To run it yourself you need Python, command lines and an NVIDIA graphics card. The README's comparison with Suno is about an internal 7B model that is not released, so what you can run today is the smaller 3B.
HeartMuLa is free and open source (Apache-2.0) and can replace Suno and Udio.
How easy is it to install HeartMuLa?
A · download and open · Web
- Windows D
- Linux D
- Web A
The easy way in is the team's hosted demo on Hugging Face and ModelScope, which needs nothing installed, so web is an A. Running it yourself is command line only on Windows and Linux: git clone, `pip install -e .`, then download several model checkpoints by hand and launch a Python script that expects CUDA devices, which is a D.
HeartMuLa on GitHub
- Stars
- 3,790on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +7stars in 6 days
Stars over time
Facts
- Runs on
- Windows, Linux, Web
- You need to install
- Nothing extra
- To build from source
- Python, Graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- None
- Cost
- Free · Apache-2.0 licence, no account or payment. The model files are downloaded free from Hugging Face or ModelScope.
- License
- Apache-2.0 (Permissive)
- Last change on GitHub
- Oct 8, 2026
HeartMuLa risk signals (4)
- Attentionneeds a strong nvidia card
The README's FAQ is titled "CUDA out of memory?" and suggests splitting the model across two cards ("e.g. 2 4090s") with `--mula_device cuda:0 --codec_device cuda:1`, or `--lazy_load true` on a single card. The scripts default to `cuda`, so there is no documented path for Mac or for a computer without an NVIDIA card.
AI review
- Infoslow generation
TODO list: "Release scripts for inference acceleration and streaming inference. The current inference speed is around RTF ≈ 1.0" — roughly one minute of computing for each minute of music.
AI review
- Infoheadline result not released
Highlight section: "Our latest internal version of HeartMuLa-7B achieves comparable performance with Suno", while the parameter list says of the 7B option "# `7B` version not released yet." The repository description also calls it "The Most Powerful Open-Source Music Generation Model of 2026".
AI review
- Infoearly stage packaging
No GitHub releases at all, pyproject.toml says version = "0.1.0" with description "A Python Library.", and install is `git clone` plus `pip install -e .` from the main branch, so you always get whatever is current.
AI review
Evidence is shown in its original language.
What Radar could not check about HeartMuLa
- How good the songs sound compared with the demo page
- Real disk space after downloading the HeartMuLa and HeartCodec checkpoints
- How much video memory a single card actually needs, since the README gives no number
- Whether the hosted demo pages stay available and how long the queue is
- Who owns the rights to the music you generate, since the repository only states the Apache-2.0 licence of the code and weights