YuE2 Guide: VRAM Requirements, ComfyUI Paths

September 25, 2026September 28, 2026

About This Model

YuE2 is a music generation model released by m-a-p (Multimodal Art Projection). Give it lyrics and a text “style" (genre, instruments, vocal character) and it produces a complete song with vocals and accompaniment as 48 kHz stereo audio. It succeeds YuE, released in 2025; the released weights are YuE2-3B, with 3.6B parameters.

Its defining trait is that it does not jump straight to audio: it first writes a score—melody and chords in ABC notation—and then renders that score as sound. You can export the score and edit it by hand, or, as the publisher documents, ask an AI agent to “reharmonize this as jazz" or “add a saxophone solo" and let YuE2 render the revision. It also supports covers that re-style an existing song; for those, the team’s transcription model SheetSage2 extracts a score from the original recording.

What Makes It Stand Out

The publisher claims song quality “competitive with Suno v5/v6," and on WildSongBench—an automatic evaluation over 192 prompts—its numbers do sit alongside commercial services.

Three things stand out in this table:

• It is clearly ahead among open models. It beats the runner-up, LeVo 2 (6.32), by a wide margin, and even its standard setting (6.73) scores above the commercial Suno v6 (6.56).
• The jump from the previous model is large. YuE scored 4.92 overall with a 36% phoneme error rate—its lyrics were hard to make out. YuE2 brings that down to about 8%, on par with Suno.
• YuE2’s figures come from generating several songs and keeping the better one. Even the standard setting picks the better of two candidates by lyric clarity, and “best of 8" keeps the highest-rated of eight. A single run will not always reach that quality. All of these are automatic metrics, not human listening tests (the publisher is collecting human preferences in its “Music Arena").

For covers, it scores 71.3% on a metric for how recognizable the original song remains (CLEWS Hit@1), far ahead of SongEcho (48.4%) and ACE-Step 1.5 (2.4%). The publisher measured a 3.6-minute song being generated in about 71 seconds on an RTX 4090.

Running It Locally

• The official recommendation is a 24GB NVIDIA GPU with BF16 support and 24GB of free system RAM, on Linux. In the publisher’s measurements, peak GPU memory on an RTX 4090 was about 11 GiB (14.08 GiB at maximum length). The hardware table below is this site’s estimate from file sizes and places the model in the 12GB tier, but compared with the measured peak above, a card in that tier leaves almost no headroom.
• For ComfyUI, start with Comfy-Org/YuE2. It repackages the files for ComfyUI and ships official workflows, “YuE2: Text to Music" and “YuE2: Music Cover." sheetsage2_bf16.safetensors is the SheetSage2 transcription model used for covers.
• The license is CC BY-NC 4.0, which prohibits commercial use. Selling songs you generate or using them in monetized videos is not allowed.
• The technical report is marked “coming soon"; for now, the paper for the original YuE (arXiv:2503.08638) is the reference.

Sources: m-a-p/YuE2-3B model card and Comfy-Org/YuE2, as of 2026-09-25. Benchmark figures are the publisher’s automatic evaluations.

Our Coverage and Data

Everything Local Model Watch has published about the YuE2 family: 3 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our model family index.

At a Glance

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 3.6B parameters

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — cc-by-nc-4.0 (Commercial use prohibited): Commercial use is prohibited (NC = NonCommercial). Internal business use can also count as commercial, so avoid this license for work use.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Distributed Files (Comfy-Org/YuE2)

Weight files published in Comfy-Org/YuE2, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

The ComfyUI folder is shown only where the repository’s layout mirrors ComfyUI’s models/ directory.

Distributed Files (m-a-p/YuE2-3B)

Weight files published in m-a-p/YuE2-3B, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

Articles (the family’s own models first, then newest)

Repositories

• m-a-p/YuE2-3B
• Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4
• Comfy-Org/YuE2

Last updated 2026-09-28 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log.