[
    {
        "role": "user",
        "content": "There%20is%20an%20article%20on%20404co%20that%20says%20that%20Suno%20ripped%20and%20scraped%20youtube%20and%20other%20music%20services%20for%20all%20of%20the%20audio%20for%20it%27s%20music%20service%20for%20AI%20music%20generation%20and%20I%20myself%20do%20not%20like%20the%20fact%20that%20they%20did%20this%20as%20no%20wonder%20why%20the%20quality%20is%20so%20low%20as%20it%20is%20just%20youtube%20clips%20and%20rendars%20so%20if%20they%20did%20that%20then%20how%20did%20they%20build%20the%20model%20on%20that%20content%3F"
    },
    {
        "role": "assistant",
        "content": "Hey \u2014 those are valid concerns. What you\u2019re describing is part of an ongoing legal and ethical debate around what Suno *did*, *how* they did it, and whether it's allowed. I dug into credible sources to lay out what we *know for sure*, what\u2019s alleged, and what\u2019s still being contested. I\u2019ll also explain how this could affect perceived quality of the music.\n\n---\n\n## What Suno *says it did*\n\nAccording to Suno\u2019s own disclosures (in a document meant to comply with California Assembly Bill 2013), the model is trained on:\n\n- \u201cTens of millions of public music audio files and corresponding textual metadata \u2026 from publicly available music files and related metadata accessible on third-party websites on the open internet.\u201d ([help.suno.com](https://help.suno.com/en/articles/9709569?utm_source=openai))  \n- They say they do not bypass paywalls, or access password-protected content. ([help.suno.com](https://help.suno.com/en/articles/9709569?utm_source=openai))  \n- Their training data includes works that may be under copyright and some that might not be; and they also use \u201cuser content\u201d and \u201cuser activity information.\u201d ([help.suno.com](https://help.suno.com/en/articles/9709569?utm_source=openai))\n\nIn court, Suno has admitted that copyrighted songs \u201cwere presumably included\u201d among the tens of millions of recordings used to train the model. ([techcrunch.com](https://techcrunch.com/2024/08/01/ai-music-startup-suno-response-riaa-lawsuit/?utm_source=openai))\n\n---\n\n## What\u2019s been alleged / uncovered in external reports\n\nHere\u2019s where the more controversial / possibly problematic stuff comes in:\n\n- A **404 Media report** (cited by other music news outlets) claims that hacked source code from Suno revealed that they scraped audio from YouTube Music, Deezer, Genius, RSS podcast feeds, stock music libraries, etc. ([tech.yahoo.com](https://tech.yahoo.com/cybersecurity/articles/hack-suggests-ai-music-generator-170034642.html?utm_source=openai))  \n- That code allegedly shows large volumes of audio ingestion from these services, including something like 2,013,545 music clips from \u201cyoutube_music,\u201d and ~113,879 hours from just that source. ([musicbusinessworldwide.com](https://www.musicbusinessworldwide.com/suno-scraped-youtube-deezer-and-genius-to-train-its-ai-hacked-code-reveals/?utm_source=openai))  \n- In legal filings, major record labels like UMG and Sony are alleging that Suno unlawfully copied (ingested) copyrighted tracks into its training dataset without permission. ([musicbusinessworldwide.com](https://www.musicbusinessworldwide.com/files/2026/05/Poseidon-v-Suno.pdf?utm_source=openai))  \n- One legal document claims Suno scraped content from YouTube in violation of YouTube\u2019s technical protection measures (TPMs), which may violate the Digital Millennium Copyright Act (DMCA). ([musicbusinessworldwide.com](https://www.musicbusinessworldwide.com/files/2025/10/Suno-new.pdf?utm_source=openai))\n\n---\n\n## What\u2019s *not* confirmed\n\n- It's not definitively proven in court yet whether the data sourcing crossed legal boundaries (though plaintiffs allege it did).\n- Suno disputes parts of the claims\u2014especially around whether what they've done qualifies legally as fair use, and whether they circumvented protections like TPMs. ([techcrunch.com](https://techcrunch.com/2024/08/01/ai-music-startup-suno-response-riaa-lawsuit/?utm_source=openai))  \n- The precise scale (total hours, all sources) of allegedly scraped data is under dispute, and Suno is trying to keep some of that under seal in court. ([musicbusinessworldwide.com](https://www.musicbusinessworldwide.com/umg-and-sony-urge-court-to-reject-sunos-bid-to-seal-the-size-of-its-ai-training-data-arguing-the-public-has-a-right-to-see-the-scale-of-its-copying/?utm_source=openai))  \n\n---\n\n## How this relates to \u201clow quality\u201d\n\nYou mentioned you think the music is low quality, potentially because it's built from \u201cYouTube clips and renders.\u201d If the videos and renders scraped were low bitrate, included distortions, compression artifacts, or had overlapping metadata issues, that could degrade the model's training signal. Essentially:\n\n- Poor\u2010quality source audio \u2192 models learn from subpar audio \u2192 generated outputs may suffer (noise, unnatural timbre, etc.)\n- If some clips are very short or heavily processed (ripped from videos), the model may fill in gaps poorly, leading to generic or \u201coff\u201d results.\n\nSo yes \u2014 if much of the training data was from \u201cYouTube clips\u201d or \u201cscraped renders,\u201d that could explain why outputs sometimes sound less polished or more \u201cartificial.\u201d\n\n---\n\n## Why Suno (and others) argue it *might* be legal / \u201callowed\u201d\n\n- Suno claims that using publicly accessible data and \u201clearning from it\u201d is fair use. They compare it to a musician listening widely across genres in the real world. ([techcrunch.com](https://techcrunch.com/2024/08/01/ai-music-startup-suno-response-riaa-lawsuit/?utm_source=openai))  \n- They assert they did *not* bypass paywalls or obtain unauthorized access\u2014at least in their public disclosure. ([help.suno.com](https://help.suno.com/en/articles/9709569?utm_source=openai))  \n- Suno also claims they perform data cleaning and filtering to remove \u201cjunk and other low-quality data.\u201d ([help.suno.com](https://help.suno.com/en/articles/9709569?utm_source=openai))\n\n---\n\n## Key Legal / Ethical Issues\n\n| Issue | What\u2019s at stake |\n|---|---|\n| **Copyright infringement** | Using copyrighted recordings without permission may violate rights. Whether it\u2019s fair use is unresolved. ([musicbusinessworldwide.com](https://www.musicbusinessworldwide.com/files/2024/06/Sunosuit.pdf?utm_source=openai)) |\n| **DMCA / TPMs** | If Suno circumvented YouTube's technical protection measures, that could be a breach of the DMCA. ([musicbusinessworldwide.com](https://www.musicbusinessworldwide.com/files/2025/10/Suno-new.pdf?utm_source=openai)) |\n| **Terms of service violations** | Scraping can violate YouTube\u2019s or other services\u2019 rules. ([tech.yahoo.com](https://tech.yahoo.com/cybersecurity/articles/hack-suggests-ai-music-generator-170034642.html?utm_source=openai)) |\n| **Quality and transparency** | Users and creators want clarity about what\u2019s used and how. Also, low-quality or derivative outputs hurt perceptions and legitimacy. |\n\n---\n\n## Bottom line\n\nWhat you\u2019re saying\u2014that Suno ripped stuff off YouTube etc.\u2014is at least *partially* backed by credible allegations (including code leaks / hacks), though not *finally proven* in court in every way. What Suno admits aligns with some of what\u2019s said: they *did* train on public, possibly copyrighted materials accessible on the internet, and they claim to clean / filter it. They argue that what they did should qualify as fair use. ([techcrunch.com](https://techcrunch.com/2024/08/01/ai-music-startup-suno-response-riaa-lawsuit/?utm_source=openai))\n\nIf you like, I can pull up the *current status* of that lawsuit to see how likely Suno is to be held legally liable, or whether they\u2019ve admitted to certain practices. Would you prefer that?"
    },
    {
        "role": "user",
        "content": "once%20all%20of%20those%20videos%20were%20downloaded%20how%20did%20Suno%20build%20the%20model%20and%20how%20large%20is%20it%20and%20what%20type%20of%20GPU%20might%20they%20use%3F"
    },
    {
        "role": "assistant",
        "content": "I couldn\u2019t find any reliable public source that gives detailed, definitive specs for *exactly* how Suno built their proprietary music model \u2014 things like model size in parameters, data volume in hours, or specific GPU hardware they used for training. Suno has not published a full technical paper with those engineering details. But based on what *is* known, informed estimates, and typical architectures in this space, we can outline plausible setups and what the constraints tell us. Here\u2019s what I discovered:\n\n---\n\n## What *is* known / publicly disclosed\n\n- Suno published a **disclosure for CA AB-2013** showing the training dataset for their *text-to-speech* branch (e.g. Bark) uses \u201cmillions of public audio files and related metadata\u201d\u2026 gathered since Fall 2022, with use starting Spring 2023. ([help.suno.com](https://help.suno.com/en/articles/9710273?utm_source=openai))  \n- For their music generation models, Suno has versioned models like **v3, v4, v4.5, v5, v5.5** which iterate on audio quality, arrangements, vocals, and other features. ([about.suno.com](https://about.suno.com/release-notes?utm_source=openai))  \n- Suno has stated that they perform data cleaning and filtering to remove \u201cjunk and low-quality data\u201d from their training collections. ([help.suno.com](https://help.suno.com/en/articles/9710273?utm_source=openai))  \n\n---\n\n## What *isn\u2019t* publicly confirmed\n\n- The **total number of training examples**, total input hours, or the size (# of parameters) of their proprietary music model (v5.5 or other) are not disclosed by Suno in technical documentation.  \n- They haven\u2019t published which GPU setups were used for full-model training (e.g., number of GPUs, GPU model, memory size, cluster specs).\n\n---\n\n## Inferred / estimated architecture & model scale\n\nUsing what is typical in large music/AU generative models + community observations:\n\n- Some analyses (e.g. on \u201cInside Suno v5.5\u201d) suggest the architecture is likely a **hybrid** of transformer-based modules (for text & style condition / lyrics generation) plus diffusion-like or auto-regressive audio generation. But these are inference/investigative articles\u2014not official papers. ([aiwiki.ai](https://aiwiki.ai/wiki/suno?utm_source=openai))  \n- Community posts (less rigorous but useful for scale guesses) suggest models in this domain might have *billions* of parameters. E.g. similar generative audio models, or open alternatives, have parameter counts in the range of **1\u201310B+** for higher-qual-music generation tasks. (Though we don\u2019t know Suno\u2019s exact number.)  \n\n---\n\n## What GPU / hardware would make sense \u2014 plausible setup\n\nGiven the scale, complexity, and latency expectations of Suno v5/v5.5, here\u2019s what a top-end plausible training/inference setup might look like:\n\n| Component | Reasoning / Requirements |\n|---|----------------------------|\n| **GPU type(s)** | High-VRAM data-center GPUs: NVIDIA A100 (40\u201380 GB), H100, or equivalent. Possibly clusters of these with parallelization (data and model parallelism). |\n| **VRAM / Memory per GPU** | Probably 40\u201380 GB+ per GPU. Large transformer blocks + upsampling / diffusion elements + audio tokenizers are memory-intensive. |\n| **Number of GPUs (training)** | Could be anywhere from a few dozen to hundreds, depending on batch size, speed target, how many training epochs, and whether they do mixed precision or other memory optimizations. |\n| **Inference hardware** | Less demanding: maybe fewer GPUs (or even powerful single GPUs) to serve real-time or near-real-time generation. They may also use optimized inference frameworks, quantization, and offloading to reduce VRAM load. |\n\n---\n\n## How they *might* build the model after scraping/downloading\n\nPutting together known and inferred pieces, here\u2019s a plausible workflow:\n\n1. **Data collection & preparation**  \n   - Collect scraped audio (from public websites, YouTube, etc. as alleged) + metadata.  \n   - Clean/filter: remove low-bitrate audio, duplicates, noisy or misaligned metadata, extremely short clips. This enhances quality.  \n\n2. **Pre-processing / encoding**  \n   - Convert audio into representations suitable for modeling (e.g. spectrograms, mel-spectrograms, audio tokens, or continuous waveform modeling).  \n   - Possibly extract features about genre, style, instrumentation via metadata or auxiliary classifiers.  \n\n3. **Model architecture**  \n   - A conditional model that takes text prompts (genre, mood, lyrics) \u2192 encoded into embeddings.  \n   - Diffusion or autoregressive model to generate audio waveform or audio token / spectrogram, followed by upsampling / vocoder to raw audio.  \n   - Possibly separate modules for lyrics generation, melody/instrumentation, vocals, mixing.  \n\n4. **Training**  \n   - Large scale training across many GPUs, with mixed precision (FP16 / BF16) to save VRAM and speed up.  \n   - Likely uses long training runs with substantial compute, regularization, music-specific augmentations.  \n   - Monitoring loss metrics for different audio quality aspects, prompt adherence, etc.\n\n5. **Inference serving**  \n   - Once model is trained, deployed for real-time or near real-time generation: likely sharded across inference servers, using GPU clusters.  \n   - Might use lower bit quantization for serving to improve latency. Possibly smaller distilled versions of full models for \u201cfree\u201d features or fast preview generations.  \n\n---\n\n## Rough estimate: how large *could* Suno\u2019s model actually be\n\n- If Suno v5.5 is their latest most-capable model, it\u2019s plausible that it has **several billion parameters**. Maybe in the range of **5\u201320 B** depending on model complexity (text + audio + vocoder + style conditioning).  \n- Given training started spring 2023 and public users by 2024, they had multiple version iterations (v3, v4, etc.), suggesting incrementally larger or improved models.\n\n---\n\nIf you want, I can try digging up any leaks, academic studies, or related models to get a narrower ballpark (parameter count + GPUs + overall compute) for Suno\u2019s v5.5, to compare it to similar systems."
    },
    {
        "role": "user",
        "content": "yes"
    },
    {
        "role": "assistant",
        "content": "I searched through public sources and leaked discussions and here\u2019s what I found about Suno\u2019s model size, compute, and GPU hardware\u2014what is known, what\u2019s speculative, and where evidence is thin. Most of this comes from community leaks and inferences (so take it with caution).\n\n---\n\n## What *is* known / confirmed\n\n- According to Suno\u2019s **CA AB-2013 disclosure**, their music generation models are trained on **\u201ctens of millions\u201d** of public audio tracks and associated metadata. ([help.suno.com](https://help.suno.com/en/articles/9709569?utm_source=openai))  \n- Suno\u2019s newer model v5.5 (launched March 26, 2026) claims **richer arrangements**, sharper vocals, and more dynamic sound versus earlier versions. Also features like *Voices*, custom models, and \u201cMy Taste\u201d personalization. ([help.suno.com](https://help.suno.com/en/articles/11362305?utm_source=openai))  \n\nThey do *not* disclose:\n\n- Exact parameter count\n- Number of training hours\n- Detailed hardware specifications (type/number of GPUs used for training)\n\n---\n\n## What the leaks / external reports suggest\n\n- A **hack revealed** datasets allegedly used, with hourly estimates for different sources. Some claims include ~113,879 hours from YouTube Music, ~17,615 hours from Genius, 12,287 hours from Deezer, etc. However, these are unverified and based on leaked internal files. ([reddit.com](https://www.reddit.com/r/BetterOffline/comments/1uxbolq/hack_reveals_suno_ai_music_generator_scraped/?utm_source=openai))  \n- In legal filings (e.g. in case 1:24-cv-11611), Suno admitted that its training data included \u201cmillions\u201d of copyrighted sound recordings owned by plaintiffs, but they\u2019ve resisted listing specific works. ([musicbusinessworldwide.com](https://www.musicbusinessworldwide.com/files/2026/05/Memorandum-SUno-1-2.pdf?utm_source=openai))  \n\n---\n\n## What\u2019s more *speculative* or estimated (informed guesses)\n\n- Some Reddit discussion (not official) suggests for *full song generation up to 5 minutes*, **80 GB of GPU RAM** might be necessary. ([reddit.com](https://www.reddit.com/r/SunoAI/comments/1ibzr9n/we_now_have_suno_ai_at_home_with_this_new_local/?utm_source=openai))  \n- The \u201cYuE\u201d model mentioned in these posts is claimed to have ~**7 billion parameters**, though it\u2019s a different model and not officially part of Suno\u2019s suite. Still, it gives inference about scale and what users expect from high-end local variants. ([reddit.com](https://www.reddit.com/r/SunoAI/comments/1ibzr9n/we_now_have_suno_ai_at_home_with_this_new_local/?utm_source=openai))  \n\n---\n\n## Comparing with similar academic models\n\nTo put things in context:\n\n- **HeartMuLa**, an open-source \u201cSuno-level, commercial-grade\u201d music foundation model, has variants with **7B parameters**. It\u2019s often cited as a baseline that approximates commercial systems. ([arxiv.org](https://arxiv.org/abs/2601.10547?utm_source=openai))  \n- In another recent project, **Muse** uses synthetic data and smaller model size but still competes in aesthetic quality and textual fidelity. So good performance is possible even with moderately sized models. ([arxiv.org](https://arxiv.org/abs/2601.03973?utm_source=openai))  \n\n---\n\n## My best guess: What Suno *probably* uses\n\nGiven all this, here\u2019s a synthesis / likely scenario for Suno v5.5 or similarly capable commercial models:\n\n- **Parameter count**: Probably somewhere between **5\u201320 billion parameters**. The architecture is likely modular\u2014text/lyrics embedding, style/genre/voice conditioning, core audio generation (diffusion / autoregressive / token-based), and a vocoder/synth module for polish.\n\n- **Training data volume**: Likely **hundreds of thousands of hours** of audio, given the leaked numbers for some sources plus \u201ctens of millions\u201d of files. Because audio is heavy data, they must have used aggressive filtering and quality control to limit unnecessary bulk.  \n\n- **Hardware setup**: \n  - Major training runs likely run on clusters of high-end GPUs\u2014NVIDIA A100 / H100 or equivalents\u2014with 40\u201380+ GB VRAM each.\n  - Probably dozens to perhaps over a hundred GPUs when doing the largest-scale training (discarding or averaging model snapshots, tuning, etc.).\n  - Use of mixed precision (FP16 / BF16) to reduce memory and improve speed, plus model/data parallelism across multiple GPUs or nodes.\n\n---\n\nIf Suno ever publishes more detailed tech specs\u2014or if there\u2019s a more confirmed leak\u2014those guesses could be refined. Want me to try to pull up any partial engineering interviews or patent filings that might narrow parameter\u2010count/GPU usage further?"
    }
]