How to Run YuE2 Locally: Hardware, VRAM and Setup

Quick answer: YuE2-3B requires an NVIDIA GPU with BF16 support and 24 GB of VRAM, running Linux with Python 3.12. It outputs 48 kHz stereo without quantization. A 24 GB floor means an RTX 3090, 4090, 5090 or workstation card — a 3060, 4060 or 4070 cannot run the official weights.

YuE2 is the strongest open-weights music model currently published, and the most common reason people cannot use it has nothing to do with quality. It is the memory requirement.

The stated requirements

These are M·A·P's own figures. You will find lower numbers quoted around the web, including memory readings taken from the older YuE 1 release; they do not describe this model.

Which GPUs actually clear the bar

GPUVRAMRuns official weightsRTX 306012 GBNoRTX 4060 Ti8 or 16 GBNoRTX 407012 GBNoRTX 408016 GBNoRTX 309024 GBYes — at the floorRTX 409024 GBYes — at the floorRTX 509032 GBYes, with headroom

This is the single biggest practical obstacle to running YuE2. A 24 GB requirement excludes most of the consumer cards people actually own.

Can you run it on less?

Not with the official weights. 24 GB is the stated floor and there is no official lower-memory configuration.

Community GGUF quantisations appear quickly for models like this, and quantisation generally lowers VRAM at some cost to output quality, so one may bring the model within reach of a 16 GB card. Treat their specifications as unverified: they are community work rather than official releases, published figures vary between builds, and none is maintained by the original team. A quantised build also will not reproduce the published benchmark scores, which were run unquantized with a specific decoder. If you intend to depend on one, test it yourself before building a workflow around it.

Apple Silicon and Windows

Neither is officially supported. The stated requirements are Linux with Python 3.12 and an NVIDIA GPU with BF16, which rules out Macs entirely and leaves Windows users on WSL2 or a dual boot.

Community ports for Apple Silicon do tend to appear for popular models. Expect to debug them yourself; M·A·P maintain none of them, and a port that runs is not the same as a port that reproduces the published quality.

Installing it

1. Clone the official YuE repository from M·A·P on GitHub.
2. Create a Python 3.12 environment — earlier versions are not supported, and this is the most common cause of dependency failures.
3. Install the dependencies listed in the repository.
4. Confirm your GPU reports BF16 support before the first run. This is the most common cause of a failed start.
5. Run the pipeline; the YuE2-3B weights download from Hugging Face automatically on first use.

The inference code is Apache-2.0. The weights are not, and the difference matters the moment anything you are building touches revenue.

What YuE2 does that other local models do not

Three capabilities separate it from the usual prompt-in, audio-out design, and they change how you work rather than just how the output sounds.

• Symbolic planning. The model writes a melody-and-chord plan that is readable, playable and changeable before any audio is rendered. The structure is a representation you can inspect, not an opaque render.
• Agentic editing. Because a score exists, you can hold a conversation about the arrangement and revise one section rather than regenerating the whole track and hoping the parts you liked survive. Anyone who has lost a good chorus to a re-roll will recognise the problem.
• Zero-shot covers. Reinterpret an existing piece in a different style with no cover-specific fine-tuning. M·A·P measure this at 0.647 CLEWS mAP with a full score against 0.006 without one, across 948 works.

None of this is captured by a benchmark score, and for day-to-day work it may matter more than a 0.2 difference in an automatic average. It is also the clearest argument for YuE2 over the open models scoring close behind it.

Why run a music model locally at all

• Your audio never leaves your machine. Nothing is uploaded, which matters for unreleased work and client material.
• No credit meter. The marginal cost of the ninth take is electricity rather than a subscription tier — which is exactly what best-of-8 sampling demands.
• You set the policy. A local model applies whatever content rules you decide, not whatever a cloud provider's terms permit.
• It keeps working. A downloaded model does not change under you, get deprecated, or raise its price.

What local actually costs

The honest comparison is not free versus paid; it is capital versus subscription.

Over 12 monthsLocal (YuE2)Cloud subscriptionUp-front24 GB GPUNonePer generationElectricityCreditsBest-of-8 samplingTime only8× the creditsCommercial rightsNone — CC BY-NCPer provider termsWorks offlineYesNo

If you generate a handful of songs a month, a subscription is cheaper by a wide margin. The economics turn toward local at volume, and turn decisively once best-of-8 is involved — eight generations per usable track burns credits quickly, but on your own hardware it costs only time.

Common problems

• Out-of-memory errors on a 24 GB card. 24 GB is the floor, not comfortable headroom. Close everything else using the GPU, including your desktop compositor if you can.
• BF16 unsupported. Older cards lack it and the requirement is not negotiable. Check before you install anything.
• Python version mismatch. 3.12 specifically.
• Quantised builds behaving differently. Community GGUF builds are not the official model and will not reproduce the published benchmarks.
• Expecting benchmark quality from one generation. Both published YuE2 scores involve candidate selection — two generations for the standard figure, eight for best-of-8. A single unselected run is not what was measured.

Is it worth it?

Be honest about which situation you are in.

• You already own a 24 GB card. Yes, without much hesitation. The model is free to download, generation is unlimited, and the capability is genuinely at the frontier of open weights.
• You would be buying hardware for this. Think carefully. A 24 GB GPU is a significant purchase, the licence forbids earning anything back from the output, and the field is moving fast enough that today's requirements may look different in six months.
• You need commercial rights. This is not your model regardless of hardware, and no amount of VRAM changes that.

The weights are CC BY-NC 4.0 — non-commercial only, and the restriction is intended to cover the music you generate, not just the model files.

Sources: M·A·P YuE repository, its docs/benchmarks.md and THIRD_PARTY_NOTICES.md, and the YuE2-3B model card. Figures checked 15 September 2026 against the WildSongBench run published 12 September 2026.