Summary
Fahd Mirza installs and live-tests YuE2, an open-weights music generation model whose benchmarks claim to outperform Suno V6 on both song quality and prompt adherence — despite running as a 3-billion parameter model on local hardware. The video covers architecture, real VRAM measurements, and generated audio samples across Mandarin, English, and multi-language outputs.
YuE2’s pipeline is architecturally interesting for its editable intermediate step: rather than generating audio end-to-end, it first produces a symbolic score in ABC notation (literal sheet music), converts that to semantic tokens via an autoregressive expert model, then uses flow matching to produce acoustic latents, which a VAE decoder renders to a 48kHz waveform. This means the intermediate score can be opened and edited before regenerating — a meaningful workflow advantage over models that treat generation as a black box.
Running on Ubuntu with a 48GB Nvidia GPU, Mirza measures just under 8GB of VRAM end-to-end, generating 212 seconds of audio in approximately two minutes. He tests the default Mandarin example from the model card, generates a custom English song from original lyrics, and probes multi-language support. His honest assessment: the instrumental quality and chorus are competitive, but the AI character in vocals remains detectable. Includes cover generation and agentic editing modes. A practical evaluation for developers and creators tracking open-source alternatives to Suno and Udio.
📺 Source: Fahd Mirza · Published September 12, 2026
🏷️ Format: Hands On Build
