This Shouldn’t Be Possible… Open Source AI Music (SUNO LEVEL)

Open-source AI music generation can now run locally on a typical gaming PC—producing multi-minute songs with lyrics and instrumentation without relying on cloud APIs. The standout model highlighted is Heart Moola, an open-weight, multimodal LLM-based system for song generation that the creator claims performs near the top of the field on lyric clarity while remaining feasible to run on consumer hardware.

Heart Moola is presented as a practical alternative to closed, top-performing music generators such as Sunno V4.5 and Sunno V5. In side-by-side listening tests, the closed models are described as “all-encompassing” and strong at both lyrics and overall musical feel, while Heart Moola—despite being smaller (about 3B parameters)—is credited with delivering surprisingly comparable results, especially when the goal is lyric-focused output. The model’s architecture is broken down into components including a text tokenizer, an audio encoder, a codec tokenizer, and a local decoder that produces generated music. The project is also positioned as “open-source weights and all,” with an Apache 2.0 license and a GitHub repository that includes code and a tutorial.

Beyond raw quality, the practical message is how to make it work offline. The walkthrough centers on using Google’s “Anti-Gravity” tool to install and configure the Heart Moola repo quickly. System requirements are emphasized: an NVIDIA GPU using CUDA, with roughly 10–12GB VRAM as a minimum and 16GB+ recommended for smooth performance. CPU-only setups are described as possible but extremely slow. On Windows, VRAM can be checked via Task Manager. The creator also notes that some setups may require “lazy loading,” and that model loading can be slower than generation itself.

In live tests, the model is run to generate songs in seconds once loaded, including attempts with missing lyrics. One experiment suggests the model can still produce vocal content even when lyrics are intended to be blank, but the creator attributes unexpected results to whether the lyrics file was actually saved and used. Another key operational detail is that the system may unload checkpoints automatically, forcing a reload that adds time; the workaround is simply to rerun and continue.

Heart Moola is also described as multilingual, with demonstrated capability in Chinese as well as Japanese, Korean, and Spanish. The project roadmap includes improvements such as inference acceleration scripts and streaming inference for web applications, plus reference-audio conditioning for finer control over generated music. A larger 7B variant is also planned, expected to require more VRAM but to move closer to the quality of full-scale closed models.

Overall, the central takeaway is that open-weight music generation is no longer just a research curiosity. With the right GPU and a guided setup flow, users can generate lyric-driven songs locally, iterate on prompts and tags, and look forward to upcoming upgrades that could make the experience more controllable and production-ready—potentially becoming the “stable diffusion of AI music.”