How to Run YuE2 Locally for Free: 8 GB GPUs, Macs and ComfyUI Now Supported

When M-A-P released YuE2-3B in early September, it was the most capable open song generator we had covered, and it came with a catch: Linux only, a 24 GB NVIDIA card, and no ComfyUI node. Three weeks later, that has changed. You can now run YuE2 locally on an 8 GB GPU through quantized GGUF files, on an Apple Silicon Mac through a native MLX port, and inside ComfyUI with both a core integration and a community node pack that adds a piano-roll editor and LoRA support.

None of this is a new model. It is the same 3.6B-parameter lyrics-to-song system we wrote up in our YuE2-3B launch post. What changed is who can actually run it, and for a working creator that matters more than the launch headline. Here is what landed, which path fits your machine, and where the limits still are.

At a Glance

• What it is: YuE2-3B, an open model that writes full songs (vocals plus backing, 48 kHz stereo, up to about 3.6 minutes) from a style prompt and lyrics, with an editable melody/chord score in between.
• What’s new: GGUF builds for audio.cpp (2.31–7.26 GB, as low as 7.8 GB peak VRAM), an MLX port for M-series Macs (4.2 GB recommended variant), native ComfyUI support and a community node pack with LoRA loading.
• What’s free: everything here. Weights, ports, nodes, and a ZeroGPU demo Space that needs no install.
• What it costs: $0 for personal use. There is no paid tier to buy.
• The catch: the weights are CC BY-NC 4.0. You cannot use the output commercially, so it is a sketchpad and demo tool, not a release pipeline.

What Changed Since the YuE2 Launch

At launch, the official model card asked for a 24 GB BF16-capable NVIDIA GPU and 24 GB of system RAM, with a measured peak around 11.2 GiB on an RTX 4090. There was no documented Windows or macOS path, the GGUF repo was marked as still in development, and the only ComfyUI repo was a placeholder.

Since then, four things shipped:

• audio.cpp GGUF support merged into main. The audio-cpp/Yue2-3B-GGUF repo now ships BF16, Q8_0 and Q4_0 weights plus a 265 MB F16 VAE. LoRA support arrived in audio.cpp 0.8.1 (September 17), and the current 0.9.0 release (September 30) has prebuilt binaries for Windows, Linux (CUDA and Vulkan) and macOS (Metal).
• A native MLX port. ahmadw/YuE2-3B-MLX runs on Apple Silicon with no PyTorch, with the VAE decoder folded in.
• Native ComfyUI support. Comfy-Org repackaged the weights as Comfy-Org/YuE2 and ComfyUI core gained YuE2 nodes, with “Text to Music” and “Music Cover” workflow templates.
• A full-featured community node pack. pytraveler/YuE2-ComfyUI (v0.11.0, Apache-2.0 code) adds a piano-roll score editor, an Edit Track node for retakes, covers from audio or MIDI, a lyrics writer and a LoRA loader.

How Good Is YuE2, Really?

Worth restating before you spend an afternoon on setup: every quality number for YuE2 comes from its makers. On their WildSongBench (192 prompts), a single YuE2 generation scores 6.73 average versus 6.87 for Suno v5 and 6.56 for Suno v6. The 6.96 headline score that “beats Suno v5” is best-of-eight, meaning you generate eight takes and keep the best. No independent leaderboard has verified it yet.

In plain terms: one take is in the same neighborhood as Suno’s recent models, sometimes behind. Budget for several takes per song. The good news is that the new ports make takes cheap. The audio.cpp maintainers measured about 39–44 seconds of wall time for a 225-second song on an RTX 5090, roughly five times faster than real time.

Which Version Should You Run?

Pick by hardware, not by hype. The file sizes and VRAM figures below come from each repo’s own measurements.

• 8–12 GB NVIDIA card: audio.cpp with Q4_0 (2.31 GB, 7.8 GB peak VRAM) or Q8_0 (4.26 GB, 8.9 GB peak). Both used the F16 VAE in those tests. Q8_0 is the safer quality pick if it fits.
• Apple Silicon Mac: the MLX port’s 8-bit folder (4.2 GB). The author calls it near-identical to BF16 at roughly twice the decode speed, and says a three-minute song takes “a few minutes end to end.”
• ComfyUI user who wants control: the pytraveler node pack. It lists 5–11.5 GB VRAM depending on song length, a 7.26 GB base model or a 3.69 GB INT8 option, and a CPU fallback that takes about an hour per song.
• No GPU at all: the free ZeroGPU Space (link below), with queue times and per-run limits.

If audio.cpp sounds familiar, it is the same no-Python engine from our audio.cpp 0.7.2 write-up, which also runs the MiniMax Music 3 GGUF build. One install covers both music models.

YuE2 vs. the Alternatives

Model	Free path	Min. hardware	Max length	Commercial use?
YuE2-3B (GGUF / MLX)	Open weights, ZeroGPU Space	~7.8 GB VRAM (Q4) or M-series Mac	~3.6 min	No (CC BY-NC 4.0)
ACE-Step 1.5	Open weights, demo Space	Under 4 GB VRAM	Up to 10 min	Yes (MIT)
MiniMax Music 3 GGUF	Open weights	~9.8 GB VRAM	~5 min	Yes, with conditions (MiniMax Community License, AI disclosure)
Suno	Free daily credits	Browser	Several min	Free tier: no. Paid plans from $10/mo: yes

If you need to sell or license the result, ACE-Step 1.5 is still the practical open choice. YuE2’s edge is control: the editable score layer, and now LoRA styles. For a different kind of control, conditioning on your own beat or melody, see DiffSynth-Music.

Try It Free: Four Ways to Run YuE2 Locally or Online

1. No install: the ZeroGPU Space

Open the YuE2-3B Space by mrfakename, type a style (“indie pop, warm female vocal, 100 BPM”), paste lyrics with section tags like [Verse] and [Chorus], and download MP3 or FLAC. A free Hugging Face login gives you more ZeroGPU time. You can also hear the lab’s curated examples on the official YuE2 demo page.

2. Windows or Linux GPU: audio.cpp + GGUF

Download the build for your system from the audio.cpp releases page (CUDA for NVIDIA, Vulkan for AMD/Intel). On a Mac, brew tap 0xShug0/audio-cpp && brew install audio-cpp. Then grab yue2-3b-q8_0.gguf and yue2-vae-f16.gguf plus the config files from the GGUF files tab into one folder, and run the model card’s command:

audiocpp_cli --task gen --family yue2 --model ./Yue2-3B-GGUF \
  --backend cuda --threads 8 --text "[verse] your lyrics here" \
  --request-option 'style=English, indie pop' --request-option cot=off \
  --session-option yue2.model_gguf=yue2-3b-q8_0.gguf

3. Apple Silicon Mac: the MLX port

pip install mlx tiktoken numpy huggingface_hub
hf download ahmadw/YuE2-3B-MLX --include "*.py" "8bit/*" --local-dir YuE2-3B-MLX
cd YuE2-3B-MLX
python generate.py --model 8bit --style "lo-fi soul, male vocal" \
  --lyrics "[verse] your lyrics here" --cot full --seed 831001 --out song.wav

4. ComfyUI: native templates or the full node pack

For the built-in route, update ComfyUI, put yue2_3b_bf16.safetensors in models/checkpoints and sheetsage2_bf16.safetensors in models/audio_encoders (both from Comfy-Org/YuE2), then open the “YuE2: Text to Music” template. For the editor, retakes and LoRAs, run git clone https://github.com/pytraveler/YuE2-ComfyUI inside ComfyUI/custom_nodes and restart. The pack includes six example workflows.

Honest Trade-Offs

• Non-commercial, everywhere. Every port inherits CC BY-NC 4.0. The Apache-2.0 license on the ComfyUI nodes covers the code, not the songs.
• Benchmarks are self-reported. The Suno comparison is the lab’s own, and the winning score needs best-of-eight sampling.
• Quantization costs something. The MLX author keeps the acoustic stage at 8-bit even in the “4-bit” build because fully 4-bit flow matching drifted. Expect Q4 GGUF to be rougher than Q8.
• Languages are narrow. English and Mandarin are the documented strengths. Other languages may work, but they are not what the model was benchmarked on.
• The native ComfyUI path has rough edges. Early coverage reported occasional ABC score failures and high VRAM during acoustic synthesis.
• Community ports are young. The MLX repo and node pack are maintained by individuals with small user bases so far. Pin a version that works for you.

Workflow Tips

1. Generate four to eight takes and pick one. That is how the headline score was achieved. With GGUF on a mid-range card, a batch of takes fits in a coffee break.
2. Lock the score, then iterate the sound. Keep the ABC output from a take you like and re-render with new seeds or styles. The melody stays put while the performance changes.
3. Use cot=off for quick drafts, full for finals. Skipping score planning is faster for testing lyrics. Turn it back on when you care about melody and harmony.
4. Try a style LoRA before writing longer prompts. Community adapters like this reggae LoRA push genre harder than adjectives do. Read each adapter’s license.
5. Demo here, release elsewhere. Use YuE2 for songwriting sketches and pitch demos, then rebuild commercial tracks with ACE-Step 1.5 or real players.

For a feel of how open and closed music models stack up by ear, watch our MiniMax Music 3.0 vs Suno 5.5 Pro head-to-head, and find more tests on the Creative AI Show YouTube channel.

FAQ

Is YuE2 free?

Yes. The weights, the GGUF and MLX ports, the ComfyUI nodes and the Hugging Face demo Space are all free. The only costs are your own hardware and electricity.

Can I use YuE2 songs commercially?

No. YuE2-3B is licensed CC BY-NC 4.0, and every port inherits that. For commercial use, look at ACE-Step 1.5 (MIT) or a paid Suno plan. This is not legal advice, so check the license text for your situation.

What are YuE2’s system requirements?

The official PyTorch release targets a 24 GB NVIDIA GPU. The audio.cpp Q4_0 build peaked at 7.8 GB VRAM in its maintainers’ tests, and the MLX 8-bit build is 4.2 GB on an Apple Silicon Mac. With no GPU, use the ZeroGPU Space.

Does YuE2 work in ComfyUI?

Yes, since mid-September. ComfyUI core has native YuE2 nodes and templates, and the community YuE2-ComfyUI pack adds score editing, retakes and LoRA loading.

Can YuE2 run on a Mac?

Yes, on Apple Silicon. Use the MLX port (8-bit recommended) or audio.cpp’s Metal build installed through Homebrew.

Accurate as of October 4, 2026. Ports, VRAM figures and release versions move fast, so check each repo before you download.