HeartMuLa
An Apache 2.0 open song model from the HeartMuLa team: lyrics and style tags in, songs of up to four minutes out.
In short. HeartMuLa writes a complete song with vocals from lyrics and a list of style tags. Its weights are Apache 2.0, so commercial use is allowed, and the README says it generates at about real-time speed.
Summary
HeartMuLa is the song generator in a family of open music models from the HeartMuLa team, which released its first checkpoint on 14 January 2026. It is a language model of about 3.9 billion parameters, named 3B, that predicts audio tokens from lyrics with section markers and a set of tags. A separate 12.5 Hz codec called HeartCodec turns the tokens into 48 kHz stereo audio, and the same project publishes HeartTranscriptor, a Whisper-based lyrics recognizer. The card lists Chinese, English, Japanese, Korean and Spanish. The paper (arXiv 2601.10547) also describes section-by-section style control and a short-clip mode for video background music, but the README lists reference-audio conditioning and fine-grained control as still to come in the open release.
The recommended checkpoint is HeartMuLa-oss-3B-happy-new-year, released on 13 February 2026, together with HeartCodec-oss-20260123. The files are stored in 32-bit precision, so the two downloads total about 22 GB. The README says generation runs at about real time (a real-time factor of about 1.0), defaults to songs of up to four minutes and suggests lazy loading when a single GPU runs out of memory. It states no minimum VRAM. The team says its larger internal 7B model matches Suno on musicality, fidelity and controllability, but that model is not released. In the YuE2 team's WildSongBench, HeartMuLa scores 6.25 on the SongBench average, behind YuE2 (6.73) and MiniMax Music 3 (6.28), and its musicality score of 5.50 is the highest of any open model in that table except YuE2.
The weights are Apache 2.0: on 20 January 2026 the team changed the license of the repository and all related weights to Apache 2.0. The September 2026 cover-song model MuLaCover, in a separate repository, is CC BY-NC 4.0 and so non-commercial. In the 30 days to 11 November 2026, Hugging Face counted about 3,100 downloads of the recommended checkpoint, far fewer than the more popular music models.
Good for
- Songs with vocals from your own lyrics and style tags, under a permissive license
- A research baseline with a paper, an audio codec and a lyrics-recognition model
- Commercial projects that need Apache 2.0 weights for a full-song generator
- Lyrics in Chinese, English, Japanese, Korean or Spanish
Watch out for
- Reference-audio conditioning and fine-grained section control, which the paper describes, are listed as still to come in the open release
- The 7B model that the team compares with Suno is internal and not released
- The checkpoints are stored in 32-bit precision, and the README states no minimum VRAM, so test on your own GPU
- The cover-song model MuLaCover is CC BY-NC 4.0, and the cards do not describe the training data
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of HeartMuLa. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| oss-3B happy-new-year with HeartCodec (FP32 files) FP32 | 3.94B | 22 GB | 26 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source music and sound-generation models in 2026
- Open music models grew up in 2026: seven models for songs, instruments and sound effects
Frequently asked questions
- What is HeartMuLa?
- HeartMuLa writes a complete song with vocals from lyrics and a list of style tags. Its weights are Apache 2.0, so commercial use is allowed, and the README says it generates at about real-time speed.
- Can I use HeartMuLa commercially?
- Yes. HeartMuLa is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is HeartMuLa open source?
- HeartMuLa is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run HeartMuLa on my own computer?
- Yes, on a good graphics card. The smallest variant, oss-3B happy-new-year with HeartCodec (FP32 files), needs about 26 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was HeartMuLa released?
- HeartMuLa was released by HeartMuLa on January 14, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- HeartMuLa-oss-3B-happy-new-year model card (Hugging Face) huggingface.co
- HeartMuLa repository: release news, license update and deployment guide (GitHub) github.com
- HeartMuLa: A Family of Open Sourced Music Foundation Models (arXiv) arxiv.org
- HeartMuLa project page and demos heartmula.github.io
- MuLaCover model card, CC BY-NC 4.0 (Hugging Face) huggingface.co
- YuE2-3B model card, WildSongBench comparison table (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| YuE2 M-A-P | 3.63B | – | CC BY-NC (Non-commercial) | 24 GB Gaming GPU | |
| MiniMax Music 3 MiniMax | 11.1B | – | MiniMax Music3 Community (Conditional) | 24 GB Gaming GPU | |
| ACE-Step 1.5 ACE-Step | 2B – 4B | – | MIT (Permissive) | 4 GB Laptop or phone | |
| Stable Audio 3 Stability AI | 460M – 1.4B | – | Stability Community (Conditional) | 4.02 GB Laptop or phone |