A song about spicy Japanese curry. That is what cocktailpeanut shared on X, under a headline that says "on your PC."
Generate SUNO-quality songs on your PC with just 4.5GB VRAM
— cocktail peanut (@cocktailpeanut) September 15, 2026
WanGP has added YuE2, the latest song generation model that's supposed to be as good as SUNO 5. And it just needs 4.5GB VRAM.
It does really feel better. Listen for yourself, a song about spicy japanese curry: https://t.co/mBzs5Sdx6b pic.twitter.com/O65WzJrRbd
The creator says WanGP has added YuE2 and that it needs just 4.5GB of VRAM 1. The headline of the post claims you can generate Suno-quality songs on your own PC 2. The creator says it really does feel better, and invites you to listen to the curry song 3. These are the poster's impressions and claims. Numbers and opinions below are attributed to their authors in the same way.
Give it lyrics and a style, get a song with vocals
WanGP's developer, deepbeepmeep, writes in the v13.00 announcement that YuE2 turns your lyrics and musical style into full songs with vocals 4.
WanGP v13.00 — It’s Your Lucky Day! 🚀
— deepbeepmeep (@deepbeepmeep) September 15, 2026
🎨 A faster, refreshed UI with 5 themes and voice dictation for prompts.
🌍 Start a generation on your PC, follow progress and manage jobs from your phone or another PC.
📊 See what’s happening at every stage, with cancellation during…
Another feature in the announcement stands out. According to the developer, you can start a generation on your PC, then follow progress and manage jobs from your phone or another PC 5. Picture it: the song renders on your desk machine while you check in from another room, like a piece of studio gear.
WanGP itself is free to use locally 6, and it offers one-click scripts for Windows and for Linux 7 8.
What makers have done with YuE2
An arrangement from a voice reference, in one shot
fnoji says they only gave YuE2 a quick, rough pass. The creator writes (in Japanese) that if one-shot output can be arranged this far, it could be quite good once refined 9.
Suno V5級らしいローカル音楽生成YuE2をクッソ雑に通してみたけど一発ポン出しでここまでアレンジ出来るなら詰めたら結構いいかもな
— えふのじ (@fnoji) September 12, 2026
大体実行VRAM 15GBぐらいで回ってた
そもそも音声参照でアレンジができる時点でMiniMax Music3より良い pic.twitter.com/qhFFiN6hhm
By the creator's account, it ran at roughly 15GB of VRAM 10. The creator also writes that being able to arrange from a voice reference already makes it better than MiniMax Music3 11. It points to a way of working where a voice reference becomes the starting point of an arrangement.
A heartbreak waltz on a weekend night
Michaelzsguo posted in Chinese. The post opens, in translation, with surprise that open-source music models have gotten this strong 12.
开源音乐模型已经强到这个地步了?!
— Michael Guo (@Michaelzsguo) October 10, 2026
YuE2已经发布一个多月了吧,差点儿错过了这个宝藏!用它做的音乐,听起来感觉比 MiniMax Music 强太多了。在 MusicArena 上和 Suno 的作品对比,我居然几乎听不出区别。
不多说了。周末的夜晚,用 YuE2 写一首心碎的怀旧歌曲《不说出口的华尔兹》。… pic.twitter.com/uTeQVTz0L9
The creator writes (in Chinese) that on a weekend night they used YuE2 to write a heartbroken, nostalgic song called "The Waltz Left Unsaid" (our translation of the Chinese title) 13. They say music made with YuE2 sounds much better than MiniMax Music 14, and that when comparing against Suno's works on MusicArena, they could hardly hear a difference 15.
"How much VRAM" depends on who is speaking
Several VRAM figures have appeared so far. They come from different conditions, so read them side by side with that in mind.
- cocktailpeanut: via WanGP, the creator says it needs just 4.5GB [[C48]]
- fnoji: the creator says it ran at roughly 15GB [[C54]]
- YuE2's official documentation: for running without quantization (BF16), the recommendation is a 24GB GPU [[F71]]
The 4.5GB figure is a poster's claim about running through WanGP, and is best treated as a separate condition from the official recommendation 1 16. Before trying on your own GPU, check which setup a number refers to.
What the official documentation says
- Supported environment: the official README lists Linux, Python 3.12, and an NVIDIA GPU with BF16 support and 24GB of VRAM [[F72]]. It also describes running it locally [[F70]]
- License: the code is Apache 2.0, and the model weights are CC BY-NC 4.0 with additional creator permission [[F68]]. Individuals, creators and musicians may monetize their outputs for free. Companies using it commercially need to contact the authors about a commercial license for the weights [[F74]]
- ComfyUI: YuE2 has native nodes and an official text-to-music workflow [[F73]]
If you want to run it on your own desk
Making a song is a loop of trial and error: fix the lyrics, change the style, generate again. For people who do not want to count the number of tries, here is what NVIDIA says about RTX Spark. The NVIDIA points below center on image and video generation; they do not mention music.
- According to NVIDIA's product page, you can generate images and video on your device as often as you like, with no subscription and no per-prompt fees [[F101]]
- On the same page, NVIDIA says that running locally gets you instant iteration, no API bills, and data that stays on your machine [[F102]]
- According to NVIDIA's blog, RTX Spark runs the full NVIDIA CUDA platform [[F104]]
A statement from ComfyUI co-founder Yannik Marek also appears on NVIDIA's blog, written in the future tense: ComfyUI users will be able to run highly complex multimodal workflows on a portable device 17.
Keep one caveat in mind: YuE2's official supported environment is Linux 18. If you are thinking about an RTX Spark machine for making music, check the official supported environment of the tool you want to run first.
Gear for when the song is ready
A generated song deserves more than a quick listen. To pick out small differences in sound, studio headphones come in handy.
If you like to sketch a melody yourself before writing lyrics and picking a style, a MIDI keyboard is handy.
And if you want to try an arrangement built from a voice reference, as fnoji did, and record that audio yourself, an audio interface makes recording easier.
Once you have a song, think video
After a song with vocals, you may want pictures to go with it.
- Making images, video and music with local models: our article on images, video and music made at home
- Handing Blender to an AI agent: our article on handing Blender to an AI agent
