YuE2 vs. MiniMax Music 3: Two Open-Weight Song Models Compared
In the space of about a month in 2026, two serious song models went open-weight. MiniMax, the Shanghai AI company behind the Hailuo video models, released Music 3.0 on August 13 with weights on Hugging Face. m-a-p followed in September with YuE2, which promptly topped its own published benchmark.
For anyone who wants to generate songs on their own hardware, that's a real choice for the first time. This comparison covers what each model does, how they score against each other, what they need to run, how their licences differ, and what happened to MiniMax's API. Facts about MiniMax come from its official blog, Hugging Face model card, licence file and API documentation as of late October 2026.
Disclosure: Song Creator Pro ships YuE2, so we have a stake. Where MiniMax Music 3 wins, we say so.
The Short Answer
- Overall quality: YuE2 leads on WildSongBench, 6.73 (standard) and 6.96 (best of 8) against 6.28. The benchmark is YuE2's own.
- Lyric accuracy: MiniMax Music 3 leads, with a 6.27% phoneme error rate against YuE2's 8.44%.
- Control and covers: YuE2 leads. It plans an editable score before rendering and is built for melody-faithful covers. MiniMax's open model has no score layer, and its cover feature lives in the API.
- Audio format: YuE2 renders 48 kHz stereo; the open MiniMax Music 3 renders 32 kHz stereo.
- Licence for self-hosting: MiniMax leads for businesses. Its community licence allows commercial use, including by companies; YuE2's official weights are CC BY-NC 4.0 with a creator permission that covers individuals monetizing their songs but not companies.
- Hosted access: Neither is great. MiniMax closed its paid music API to new users in August 2026; YuE2 has no official hosted service.
What Each Model Is
MiniMax Music 3
Music 3.0 generates a complete song (composition, arrangement, vocals and production) from a text description and optional lyrics, up to five minutes long. According to its model card, it's a hierarchical design: an 8B "global" language model initialized from Qwen3-8B predicts the coarse audio codes, a 0.6B local model fills in the rest, and a 2.4B flow-matching model plus a Flow-VAE decoder adapted from MiniMax Speech produce the audio. That puts it at roughly 11 billion parameters across components.
Control comes through section tags such as [Verse], [Chorus] and [Instrumental], plus what MiniMax calls Structured Captions: a detailed description of style, vocals, instrumentation and arrangement. MiniMax's blog also describes a prompt enhancement system that expands a short description into a detailed one. The model card is candid that tempo, key, instrumentation and structure may not always match what you asked for.
YuE2
YuE2 is a roughly 3B-parameter lyrics-to-song model from m-a-p. Its distinguishing feature is a planning stage: before rendering audio, it writes the song's melody and chords as ABC notation, which you can read, edit, or replace with your own. Planning can be set to full, melody-only, or off. Output is 48 kHz stereo, and the developers' demos run up to five minutes. For a full rundown, see our YuE2 review.
Quality: The Benchmark
The only head-to-head numbers come from WildSongBench, the 192-prompt evaluation published with YuE2. Interestingly, the YuE2 team listed MiniMax Music 3 in its open-model table, alongside its predecessor in the proprietary table.
| System | SongBench average | Musicality | Phoneme error rate |
|---|---|---|---|
| YuE2 (best of 8) | 6.96 | 6.27 | 9.79% |
| YuE2 (standard) | 6.73 | 5.91 | 8.44% |
| Suno v6 | 6.56 | 5.66 | 7.58% |
| MiniMax Music 2.6 | 6.32 | 5.44 | 24.55% |
| MiniMax Music 3 | 6.28 | 5.35 | 6.27% |
| ACE-Step 1.5 | 6.01 | 5.16 | 7.46% |
Source: YuE2 model card, September 2026.
Three things stand out:
- YuE2 wins on overall quality by a clear margin, even at the standard setting (which picks the better of two candidates) rather than best-of-8.
- MiniMax Music 3 sings the words more accurately. Its 6.27% phoneme error rate was the best of any open model in the table and better than Suno v6. If lyric intelligibility is your top priority, that matters. Its predecessor, Music 2.6, was at 24.55%, so this is a big improvement.
- The source is YuE2's authors. MiniMax published no quantitative benchmarks for Music 3, only qualitative claims of more natural vocals and clearer sound. No independent comparison exists yet.
Feature Comparison
| Feature | MiniMax Music 3 (open weights) | YuE2 |
|---|---|---|
| Released | August 13, 2026 | September 2026 |
| Size | About 11B across components | About 3B |
| Max song length | Up to 5 minutes | Demos up to 5 minutes |
| Output | 32 kHz, 16-bit stereo WAV | 48 kHz stereo |
| Section tags | Yes | Yes, followed closely due to planning |
| Editable score (melody, chords) | No | Yes (ABC notation) |
| Bring your own melody | No | Yes |
| Covers | Via the music-cover API model, not described in the open model card |
Yes, score-based, in the open model |
| Instrumentals | Yes ([Instrumental] tags; is_instrumental in the API) |
Yes, via empty section labels |
| Official language list | None published; demos in English and Mandarin | None published; official samples in English and Mandarin |
| Backend | CUDA (NVIDIA) | CUDA (NVIDIA) |
On covers specifically, YuE2's model card reports that on the SHS100K cover benchmark, its covers were identified as the correct original song 71.3% of the time. MiniMax's API has a dedicated cover model that takes 6 seconds to 6 minutes of reference audio, but it isn't part of the open release and MiniMax hasn't published comparable numbers. Our YuE2 remix guide shows the YuE2 cover workflow step by step.
What Happened to the MiniMax API
Until this summer, MiniMax's most attractive feature was price: $0.15 per song of up to five minutes through its API, plus a free rate-limited tier. That's still listed on its pricing page, but with an important change.
MiniMax's API documentation now states that from August 20, 2026, the paid Music and Lyrics Generation APIs are no longer available to new users, existing paying users can keep using them, and the free music models will be discontinued. MiniMax points everyone else to two places: its MiniMax Audio web product, which uses credit-based plans, or the open Music 3 weights on Hugging Face.
So if you're arriving now, the cheap per-song API route is effectively closed. Third-party resellers offer MiniMax Music 3 through their own APIs, at their own prices and terms. For the full picture of the hosted product, see our MiniMax Music alternatives page.
Want a local song model without the setup? Song Creator Pro runs YuE2 on Windows with NVIDIA or AMD GPUs from 8 GB of VRAM, plus ACE-Step 1.5 for 50+ languages and Intel GPUs. One-time purchase, no API keys, no credits, unlimited generations. The Microsoft Store version has a free trial.
Running Them Locally
Both models are designed for NVIDIA GPUs. Here's what the official sources say:
| MiniMax Music 3 (official) | YuE2 (official) | YuE2 in Song Creator Pro | |
|---|---|---|---|
| GPU | NVIDIA (CUDA) | NVIDIA, 24 GB, BF16 | NVIDIA or AMD, 8 GB+ |
| Full precision VRAM | Under 24 GB | About 11 GiB peak measured on an RTX 4090 | Not applicable (quantized) |
| Low-VRAM option | CPU offload (about 22 GB GPU) or layer streaming on 8 GB, slower | None official | Quantization plus offloading, built in |
| Setup | diffusers, SGLang-Omni or ComfyUI | Linux, Python 3.10+ | Windows app |
| Interface | Code or ComfyUI nodes | Code (ComfyUI nodes also available) | Desktop app |
MiniMax deserves credit here: an official 8 GB layer-streaming path is more than YuE2's official release offers. The catch is that it's slower, and like any self-hosted model you're assembling the environment yourself. Both models also have native ComfyUI support.
If you'd rather not touch Python at all, Song Creator Pro packages YuE2 into a Windows app that runs on 8 GB NVIDIA or AMD cards. It doesn't include MiniMax Music 3. Our guide to running YuE2 on an 8 GB GPU explains how it fits, and best GPUs for AI music generation covers which cards to look at.
Licensing and Rights
This is where self-hosters should pay the most attention, and it's MiniMax's clearest advantage.
MiniMax Music 3 is released under the MiniMax-Music3 Community License. It allows commercial use, modification and redistribution, with conditions:
- Commercial products must prominently display "MiniMax-Music3" in their user interface.
- If your company's yearly revenue (including affiliates) exceeds US$20 million, you need prior written authorization from MiniMax.
- An acceptable use policy applies, and if you let other people generate with it, you must keep reasonable safeguards against infringing uses.
YuE2's official weights are released under CC BY-NC 4.0 with an additional creator permission. Individual creators and musicians can self-host YuE2 and monetize what they make, for example on a monetized channel, with no fees or royalties. Companies using the weights commercially need a separate commercial licence from the team.
Song Creator Pro uses YuE2 with direct permission from its developers and includes a commercial licence; see the Terms of Service for details. For the broader question of who owns AI output, see who owns AI-generated music.
Which Should You Use?
Choose MiniMax Music 3 if you:
- Are a company that wants to self-host and use the output commercially without negotiating a licence
- Care most about lyric intelligibility
- Already have a 24 GB card, or can live with slower 8 GB layer streaming
- Are an existing MiniMax API customer with access to the hosted version
Choose YuE2 if you:
- Want the higher overall quality score on the only published head-to-head
- Want to edit a song's melody or chords instead of rerolling
- Make covers or genre swaps that keep the original melody
- Want 48 kHz output
- Want a Windows app on an 8 GB NVIDIA or AMD card, through Song Creator Pro
For the wider field, our best open-source AI music models roundup and YuE2 vs. Suno comparison fill in the rest.
Ready to run YuE2 without the setup? Get Song Creator Pro on the Microsoft Store for $49.99 (free trial included) or on itch.io for $44.99. One-time purchase, YuE2 and ACE-Step 1.5, unlimited local generations on your Windows PC.
Compare Song Creator Pro
All comparisonsFrequently Asked Questions
On the WildSongBench evaluation published with YuE2, YuE2 scored a 6.73 SongBench average at its standard setting and 6.96 with best-of-8 selection, against 6.28 for MiniMax Music 3. MiniMax Music 3 had the lower phoneme error rate (6.27% versus 8.44%), meaning it sang lyrics slightly more accurately. The benchmark was run by the YuE2 team, so treat it as indicative rather than final.
Its weights are published on Hugging Face (MiniMaxAI/MiniMax-Music3) under the MiniMax-Music3 Community License, which permits commercial use but requires showing 'MiniMax-Music3' in your product's interface and written authorization from MiniMax if your yearly revenue exceeds US$20 million.
Only if you were already a paying user. MiniMax's API documentation says that from August 20, 2026, the paid Music and Lyrics Generation APIs are no longer available to new users, and the free music models are being discontinued. MiniMax points new users to its MiniMax Audio web product or the open weights.
MiniMax's model card says full precision fits under 24 GB of VRAM, CPU offloading uses about 22 GB, and a slower layer-streaming mode runs on 8 GB cards. Both need CUDA (NVIDIA). Official YuE2 asks for a 24 GB NVIDIA GPU on Linux; Song Creator Pro runs YuE2 on Windows with 8 GB NVIDIA or AMD GPUs.
MiniMax has not published an official language list for Music 3; its official demos use English and Mandarin. YuE2 sings in English and Mandarin, the languages its developers' official samples cover. For other languages, Song Creator Pro's ACE-Step 1.5 model handles 50+.