HeartMuLa brings Suno-level music generation to open source

HeartMuLa just landed as a brand-new open-source music foundation model that turns text prompts and lyrics into fully finished songs with clean vocals and backing tracks.

Instead of piecing together fragments through separate tools, you feed it a genre description and lyrics, and it synthesizes the whole track. The team has already released the 3B parameter weights, which take up about 16GB of storage, with code and docs up on GitHub and Hugging Face alongside an arXiv paper.

How it actually works

Full-song generation has historically been a headache for open-source AI, usually ending up with muddy audio or forgotten melodies halfway through. HeartMuLa attacks this with four interconnected systems.

HeartCLAP handles text-audio alignment, while HeartTranscriptor takes care of robust lyric parsing. The secret sauce is HeartCodec, an ultralow-frame-rate 12.5 Hz audio codec tokenizer that preserves acoustic detail while letting an autoregressive language model manage long-form musical structure.

The main HeartMuLa model ties these together. It supports fine-grained section control, letting you prompt specific vibes for the intro, verse, or chorus, plus a dedicated preset tailored for short-form background tracks.

The good and the rough edges

In real-world tests, HeartMuLa stands out for structural consistency and audio clarity. It actually remembers motifs across different song sections instead of drifting into an entirely different track after the first verse.

It handles English, Chinese, Japanese, Korean, and Spanish. Benchmark numbers from the paper claim its lyric error rate actually beats closed commercial systems like Suno V5, while coming remarkably close in overall acoustic fidelity.

That said, it still has quirks. If your lyrics are wildly asymmetric, the model tends to drag the chorus out endlessly and eventually lose its musical train of thought. Neat, predictable verse-chorus formatting keeps it happy.

You will need a decent GPU to comfortably host the 16GB weights of this 3B model, and an even beefier 7B checkpoint is slated for release soon.

Why it matters

For indie creators, game devs, and researchers, HeartMuLa is proof that Suno-level music generation is no longer locked behind proprietary API paywalls.

Academic-scale compute and data can now reproduce commercial-grade song generation. We think this opens up genuine local music workflows where you can fine-tune your own instruments, customize arrangement styles, and never worry about platform rate limits. If you have the VRAM to spare, it is definitely worth pulling from GitHub today.