The Suno Stack
Public sources, not a leak

Suno is a Next.js music product wrapped around a closed token-transformer.

Cloning the site is easy. Cloning the sound is a research lab, a catalog, and a legal department. In 2026 you can get surprisingly close with open weights — still not v5.

Product
Next.js, React, TypeScript on Vercel
Public model
Bark: 3×80M GPTs + EnCodec
Secret sauce
Chirp v2–v5.5. Closed weights.

The stack they will admit to

Frontend is not a rumor: Suno hires for TypeScript / React / Next.js, and Vercel lists them as a Next.js + Functions customer. The music model is the rumor. Bark is the only architecture they fully published.

Client

Next.js, React, TypeScript, CSS. Job posts ask for Storybook, Sass/vanilla-extract, WCAG, and fast audio UX.

Edge / app

Hosted on Vercel. Vercel Functions for server work. Product also ships iOS and sits inside Microsoft Copilot.

Generation

Closed PyTorch models in the Bark/Chirp family: hierarchical transformers over residual vector-quantized audio tokens, plus a language model for lyrics and structure.

Serving

GPU inference with the usual production bag: distillation, mixed precision, speculative decoding, sharding. Target is a full song in tens of seconds.

Product around the model

Credits, library, feed, Extend / Cover / Persona, Add Vocals, custom models, stems, Studio editor, Songkick after the WMG deal.

How the audio actually happens

Suno does not write MIDI and then render a soundfont. It treats audio like language: compress the waveform into discrete codec tokens, then train GPT-style models to predict the next token. Bark, released April 2023, is the diagram they let everyone see.

Causal GPT80M

Plans what should be heard: timing, cadence, rough content. Not the waveform yet.

In
BERT-tokenized text
Out
Semantic tokens (vocab 10,000)

This cascade is public because Suno released Bark in 2023. Chirp, the production music stack, is the same family of idea at a scale they have not published: tokens, transformers, codec. Not a DAW. Not a concatenative sample pack.

Bark is not the product. Chirp is.

Bark is a text-to-audio toy that can speak, laugh, and occasionally sing. The consumer app generates structured songs with lyrics, vocals, arrangement, extend, cover, and personas. Internally that line is Chirp. Same token-transformer family, unpublished scale, plus an LLM for lyrics and form.

  1. BarkApr 2023Open text-to-audio. Three 80M transformers + EnCodec. Speech, SFX, crude music. This is the public architecture.
  2. v2Fall 2023Consumer music. ~1:20 generations. Discord, then suno.com.
  3. v3 / v3.520242 min, then 4 min first-gen with better song form. Extend added.
  4. v4Nov 2024Clearer vocals. Cover and Persona. Still closed weights.
  5. v4.5 / v4.5+2025Up to 8 min. Style mashups. Add Vocals / Add Instrumental as latent edits.
  6. v5 / v5.52025–26Current consumer quality bar. Voice cloning, custom models, “My Taste.” Weights remain private.

How easy is a clone?

Split the question. The website is a weekend. A credible open-source music generator is a GPU box and ACE-Step or YuE. Matching Suno v5 is not an engineering ticket — it is a company.

The website
Easy

Prompt box, library, player, credits UI. A competent frontend weekend. You are looking at a slice of that in this repo.

The product
Medium

Queues, object storage, CDN audio, social feed, moderation, billing. Normal startup work. Not the hard part.

Open-source songs
Hard, now possible

ACE-Step 1.5, YuE, MiniMax Music 3. Full songs with vocals on a consumer GPU. Quality is in the same conversation as Suno, not the same room.

Suno v5 itself
Lab + catalog + lawyers

Closed weights. Years of data and GPU time. Lyric–vocal alignment, arrangement, mix. Then the rights mess: RIAA suit, WMG settlement, BMG license. You do not clone this by reimplementing Bark.

If you actually want songs

Do not fine-tune Bark and expect a pop vocal. Plug a current open music model behind a UI like the toy studio in this repo. That is the honest 2026 clone path.

Best practical open pick. MIT/Apache-family licensing, vocals + instrumentals, runs in a few GB of VRAM, fast.
Lyric-led songs, strong vocals. Heavier (plan on 16 GB+). Closest “feels like Suno” open model if you have the GPU.
Open weights, songs up to ~5 min, 32 kHz stereo. New as of 2026; less battle-tested tooling.
Mature Meta baseline for instrumentals. CC-BY-NC on the big weights — fine for research, awkward for a product.

The one-line answer

Suno = Next.js app + lyrics LLM + hierarchical audio-token transformers + neural codec + a pile of GPU serving. The clone that matters is not a React restyle. It is whether you are willing to run someone else's 3–7B music model — or spend like a lab and license like a label.

Sources