ACE-Step 1.5 vs YuE2-3B: local music recipe
A practical split between a commercial-friendly product path and a more experimental YuE2 path with model-weight licensing limits for companies.
For a company shipping a paid local music application, licensing is part of the engineering decision, not a footnote.
AILocalabs currently prefers ACE-Step 1.5 for the commercial MusicKiln path. YuE2 is technically interesting, but its model-weight terms make it a separate research/creator recipe unless a company obtains a commercial license.
Build it yourself
- Start with ACE-Step when you want a commercial-friendly local baseline and a straightforward product path.
- Try YuE2 when its planning, cover or agentic editing workflow is the actual reason for the experiment.
- Benchmark generation time, VRAM, prompt adherence and output quality on the same prompts before switching a production backend.
- Pin the exact model and runtime revision; music stacks change quickly and a “same model name” benchmark can become meaningless.
What can go wrong
- Do not confuse an open-source code license with the model-weight license.
- Output copyright/style questions remain separate from the software license.
- A better demo is not enough reason to migrate a production product without regression and packaging tests.
What we ship — and what we do not
MusicKiln follows the ACE-Step commercial path. We publish YuE2 as a comparison/research route rather than silently bundling it.
Tried it? Leave results, questions and improvements.
Discussion stays with this recipe so future builders can learn from it. Never post passwords, API keys, license keys or private files.
Loading discussion…