
Can a music generation model make something that actually feels punk?
After experimenting with YuE2 locally on my PC, I wanted to compare it with Suno. There are a lot of benchmarks on the internet, but music cannot always be explained by numbers. Especially punk music.
The purpose was simple: push both tools towards punk, listen, adjust, and see which one captured my intention.
Something fast, aggressive, rough around the edges. Something anti-establishment.
This is a personal listening experiment. No scientific benchmark, no objective ranking of music models. Just some generations, my ears, and a question that became more interesting along the way: what exactly am I doing?
Spoiler: Having fun!
A bit about Suno
For this battle, the Suno model was v6-mini, the faster variant of its v6 family. Suno provides the model through its online service, but does not publish its weights or detailed architecture. I could give it lyrics and a style prompt, then judge the finished audio. I could not inspect the musical representations or generation steps behind it.
Suno has published an earlier audio model, Bark. Released in 2023, Bark uses three transformer stages: text becomes semantic audio tokens, then coarse audio codec tokens, then finer codec tokens. It can produce speech, sound effects, and some music. Bark shows how Suno approached generative audio at the time. Its repository explicitly distinguishes it from Suno’s text-to-music models, so I cannot use Bark’s architecture as a description of v6-mini.
In the YuE2 team’s WildSongBench, Suno v6 scored 6.55 on the SongBench average, against 6.73 for YuE2, across 192 prompts. Those figures come from automated evaluations with specified candidate-selection procedures. The benchmark did not test v6-mini, and its YuE2 results used symbolic planning, whereas my punk experiment also explored generation with planning disabled. The numbers give some context for model quality, but they cannot predict which track will sound more punk.
A bit about YuE2
YuE2 is an open-weight music model. Its generation pipeline is published, so I can describe what happens between my prompt and the finished song. Given lyrics and a style description, it can first write a symbolic plan for the melody and chords. An autoregressive transformer then generates semantic music tokens in sequence. A non-autoregressive stage turns those tokens into acoustic latents using flow matching, and a VAE decodes the latents into 48 kHz stereo audio.
The planning stage is optional. With cot=full, YuE2 plans melody and chords. With cot=melody, it plans only the melody. With cot=off, it generates without a symbolic score. That gave me a way to test whether explicit composition planning affected the punk result. I ran the YuE2-3B model in Q8 format through audio.cpp and CUDA on my RTX 4080 with 16 GB of VRAM. In these trials, disabling planning produced a rougher result to my ears.
The benchmark cited above evaluated YuE2 with melody and chord planning. Its standard setting scored 6.73 on SongBench, while a selected best-of-eight setting reached 6.96. Those figures describe a different generation setup from my cot=off punk trial. They tell me YuE2 can compete on measured song quality, but they cannot explain why one of my generations felt more punk than another.
What is punk?
For me, punk is a culture. For my AI, punk is music…
For sure, AI cannot be the judge in this battle ;-)
Even Wikipedia’s definition is confusing: https://en.wikipedia.org/wiki/Punk_rock
Make Some Noise for the Contenders!
For style, both received minimal instructions: make punk music. No BPM, chords, or other instructions.
For the lyrics, my friend GPT-6 Sol helped me. I felt that he could get more and more aggressive. Maybe it has already learned from us…
I didn’t want to make thousands of tries. Three at most. I selected the best result from each for this battle.
One useful variable on YuE2 was its CoT setting. Comparing generation with planning enabled and disabled gave me different impressions.
With full planning, the results felt more constructed. With planning disabled, I found something rougher and closer to the punk direction I wanted.
That is an observation from these trials, not proof that planning makes music less punk. Changing the mode can change the composition too. Even keeping the same seed does not turn the comparison into two identical performances with one isolated difference.
YuE2 gave me cleaner results, clearer vocals, and musical ideas I enjoyed. It sounded more organised, sometimes closer to a mainstream interpretation of the brief.
Suno sounded dirtier and more aggressive. Its metallic edge could be unpleasant in another context, but here it contributed to the feeling.
The uncomfortable part was that I preferred listening to one result while thinking the other fit the assignment better.
The winner
Suno wins the punk battle. YuE2 wins for musical quality in this experiment.
Here is how the results compared to my ears:
| Criterion | Suno | YuE2 |
|---|---|---|
| Sound | Metallic, dirty, aggressive | Cleaner, more polished |
| Musical construction | Rougher | More structured and creative |
| Listening pleasure | Less enjoyable to me | More enjoyable to me |
| Punk attitude | Stronger: raw and aggressive | Weaker: more controlled |
| Approach | Proprietary online service | Open-weight model running locally |
There is a funny contradiction here. Running a model on my own PC feels closer to the DIY spirit of punk. But the sound coming from the commercial platform was more punk.
YuE2 made music I preferred listening to. Suno captured the attitude I was asking for.
Suno made better punk. YuE2 made better music for me. And that is why Suno won this battle.
Hip-Hop Battle
Same two opponents, a different musical challenge.
Listen to both versions. What made the difference: the voice, the beat, the flow, or simply the feeling.



