Train v6 on your catalog, not on adjectives.

Do it — but build one model first, not three. Your catalog is exactly the use case Custom Models were designed for: a finished, released body of work with a recognizable sound, sitting in WAV form. Most users training a Custom Model are scraping together six demos. You have albums. That advantage is real, and it is also the trap: with 74 tracks spanning different eras, energies, and production decisions, the fastest way to waste the feature is to upload the catalog as-is and teach v6 a blurry average of everything you have ever done.

How to read the evidence labels in this guide

Suno's v6 Custom Model documentation is thin — thinner than it should be (see §02). So every load-bearing claim here carries a label. Do not treat them as equal:

A Custom Model is not a style filter and not a prompt modifier. It is a private fine-tune: a version of the model re-weighted toward your musical identity, active at the foundational level of every generation you run through it. Community-tested A Style prompt tells Suno what to do with one song. A Custom Model tells Suno who the artist is, and holds that across every song.

What comes through

• Arrangement patterns — how your tracks open, where verses sit, how you build to a chorus, bridge habits, outro behavior.
• Instrumentation fingerprint — the instruments, synth textures, drum sounds, and guitar tones you reach for repeatedly.
• Harmonic language — chord progressions, key tendencies, melodic intervals you gravitate toward.
• Energy profile — where energy rises and releases across a track; your dynamic arc as a default shape.
• Mix characteristics — EQ tendencies, stereo width, compression feel, how loud and dense your masters sit.
• Your flavor of genre — not “rock” in general, but the specific way your catalog does its lane.

This list is where Suno's Help Center description and independent testing converge. Official Community-tested

What does not come through

• Your voice. The single most misunderstood point in testing. Custom Models learn musical identity, not vocal identity — the literal timbre of a singer is not the primary carrier. That is what Voices is for. Expect a Custom Model to give you your band, your room, your arrangements — with a different singer showing up unless you also apply a Voice. Community-tested
• Specific melodies and lyrics. Suno states Custom Models capture tendencies, not memorized material — they do not reproduce your training tracks' melodies or words. Official
• Anything you did not feed it. A model trained on your downtempo work will not magically know your uptempo instincts. The model is the set.

Custom Models arrived with v5.5 on March 26, 2026 — a personalization layer (Voices, Custom Models, My Taste) built on the v5 engine. v6 launched September 9, 2026 and changed the ground underneath: Official

• Three models replace everything: flagship v6 (reliable, precise, polished), v6-wild (less predictable, more textured — for exploration), and v6-mini (fast, free for everyone). v6 and v6-wild require Pro or Premier.
• Every prior model was retired — v4 through v5.5 are gone from the picker with no rollback. Your old songs stay in your library, but any new generation, extension, or cover is a v6 generation. Prompts calibrated against 4.5+ behavior will land differently; expect to re-tune.
• Built with industry partners: v6 was developed with Warner Music Group, BMG, and Believe, with opt-in artist programs in development. Universal and Sony litigation remains open — relevant context for §10, not a reason to avoid the feature.
• New v6-native workflows that pair well with a Custom Model: plain-language section editing (fix one chorus without regenerating), mashups from multiple sources, sample/isolate-to-beat, and multimodal input (text, audio, image, video).

Plan reality (September 2026 terms)

Download caps took effect September 3, 2026: Free gets up to 7 lifetime trial downloads (no commercial rights); extra downloads are purchasable. Commercial use now attaches to permitted downloads on a paid plan — generation alone is not the license event anymore. Official

The governing finding from independent testing is blunt: coherence beats volume, every time. A tight set of 15 songs that all sound definitively like the same artist outperforms 30 songs that drift across experiments and half-finished ideas. Every track in your upload set is a vote for what the model thinks your project sounds like — and experiments dilute the signal. Community-tested

Build from decisions, not experiments. The songs that definitively represent the project. If a track makes you hesitate before adding it — it probably does not belong in the dataset.

The math of your catalog

• 74 tracks is not a training set. It is three training sets plus a reserve. At 10–15 tracks per model, your catalog can field three sharp models (30–45 tracks) and still leave you a held-out validation set and spares for rebuilds. Dumping all 74 into one model teaches v6 “the average of everything Quell has ever released” — which is a sound no single record of yours actually has.
• The era problem. If your catalog spans production eras — different mix engineers, different loudness targets, an early sound and a current sound — one model averages across them and lands in a mushy middle that matches none of them. Era is the strongest natural fault line in most catalogs, and yours is no exception.
• The outlier tax. One genre excursion in a set of twelve does not add “range.” It adds noise the model will spend identity on. Range comes from having a second model, not from contaminating the first.

Try it: the coherence × count trade-off

Drag the sliders. This is a thinking tool built from the testing consensus above — not a Suno formula (Suno publishes none) — but it will keep you honest about what a given upload set can actually teach. Expert practice

Three ways to split the catalog — pick your architecture

Do not select by favorites. Your favorite track might be your least representative one — artists often love their outliers precisely because they are outliers. Score every candidate on three axes, 1–5 each, in a spreadsheet or on paper before you open Suno: Expert practice

The decision rules

• Core set = Identity ≥4, Quality ≥4, Coherence ≥4. No exceptions for sentiment. This is the set that teaches.
• High identity, lower quality (an early signature track with a dated master): belongs in an era model if you build one — not in Core, where its mix fingerprint (thin low end, dated loudness) becomes part of what “you” means.
• High quality, low identity (the polished experiment, the genre tourism track): exclude, however good it is. It spends the model's identity budget on a sound you do not actually trade in.
• One canonical master per composition. If Drive holds a single, an album version, and a remaster of the same song, pick exactly one. Two versions of one song is a double vote for that song's specifics — it warps the set toward a single composition. (Deliberate double-weighting of a defining track is a real tactic some testers use; treat it as an advanced experiment for a rebuild, never a default.) Expert practice

Exclude on sight

• Covers. The composition is not yours. Beyond the rights problem (§10), a cover teaches the model someone else's harmonic and melodic decisions wearing your production.
• Tracks built on uncleared samples or loops you cannot document. Suno runs copyright detection on uploads; sample-built tracks are the most likely to trip it — and the least defensible if they do. Community-tested
• Collaborations where you do not control the full master — features, split-ownership beats, a guest topline you licensed for one release. See §10; when control is fuzzy, the track is out.
• Skits, intros, interludes, live recordings with crowd noise. They are votes for non-music or for a room you do not want baked into studio generations. (Exception: if a live energy is a lane you want, that is a deliberate lane model — built from several live tracks, not one stray live cut in a studio set.)
• Unfinished anything. The model cannot tell a deliberate minimal arrangement from an unfinished one. It will learn both as “your style.”

Quality vs. variety — the actual trade-off

Within your core lane, variety of songs, unity of sound. You want different tempos, keys, and song structures — that variety is what teaches the model your grammar rather than one song's shape. What you do not want is variety of production worlds. A practical check: play your candidate set on shuffle. If it sounds like one artist's album, you have a training set. If it sounds like a playlist, you have a blur waiting to happen.

Most Custom Model students upload whatever MP3 they can find. You have WAV masters. That matters more than any other prep variable, because the model learns mix and production characteristics from the files themselves — lossy artifacts and transcode mush are, literally, part of the lesson if you feed them in. Expert practice

Format rules

• Upload the WAV master (44.1 or 48 kHz, 16- or 24-bit) — the same file you'd send a distributor. Suno's guidance for training material is high-quality audio; WAV or high-bitrate MP3 is the documented expectation. Official Community-tested
• Never upload a streaming rip or an MP3 made from an MP3. If the only copy of a track is a 128 kbps MP3, that track waits for a rebuild where you have located the master — or it does not go in.
• One file per song, full song. Full forms teach arrangement arcs — intro behavior, second-verse variation, bridge placement. Do not trim tracks down to “the good parts”; the arc is the good part, to a model.

Full mixes, not stems — settle this before you touch your stem folders

Suno documents Custom Model training input as songs — finished tracks from your catalog. Stems are an output feature (Auto Split, Split from Mix, Advanced Split), not a documented training input. Official There is no published evidence that uploading stems teaches the model anything deeper — and a clear mechanism by which it hurts: each stem file counts as a “song,” so twelve stems of three tracks would teach a “catalog” of isolated drums, basses, and a cappellas with no arrangement relationships at all. Upload full mixes. Keep your stems for your DAW work, where they belong. Expert practice

Loudness and dynamics: consistency is the lesson

• Do NOT re-master or re-limit tracks for training. Your masters' dynamics — how hard they hit, how much they breathe — are part of the fingerprint the model is supposed to learn. “Fixing” them first teaches the model your mastering engineer's rescue job, not your sound.
• Do check consistency across the set. If half the set sits around −8 LUFS integrated (loud, dense, modern) and half around −16 LUFS (dynamic, older chain), the model receives contradictory instruction about how loud and dense “you” are — one more reason era-splitting beats set-mixing. Spotify normalizes playback to −14 LUFS, but that is a playback fact, not a training target: do not normalize your training files to −14.
• Do fix file-level defects only: true clipping or DC offset baked into a file, a wrong-speed or corrupted render, a file that is accidentally mono. Those are errors, not character.

Trimming and file hygiene

• Trim dead air: more than ~1 second of silence at the head or a long silent tail teaches the model your songs start with nothing. Trim to a tight 5–10 ms fade. Keep intentional intros — if the quiet build is the song, it stays.
• Remove hidden tracks and concatenated files: album files with a hidden song after silence are two lessons in one file, and the second one is a surprise. Split or exclude.
• Clean filenames before upload: Artist_Title_Master.wav beats final_v3_REAL_final2.wav when you are auditing what went into a build six weeks later. Your future model log (§09) will thank you.
• Stage a training folder in Drive: one folder per model build — CustomModel/QUELL-CORE-v1/ — containing exact copies of the uploaded files. Never upload straight from album folders; you will lose track of which master version went in.
• Verify each file plays end-to-end before upload. A truncated master teaches a truncated arc.

The build itself is the shortest part of this whole process — minutes. Everything before it (selection, prep) and after it (validation) is where the quality lives. Official Community-tested

1. Confirm the plan gateCustom Models require Pro or Premier. Free accounts cannot build one. Check you are on a paid plan before staging files, not after.
2. Open Custom Models and create a new modelIn Suno, go to Custom Models → Create New Model (in the v6 interface, watch for it in the model picker — it has been seen marked Beta Unverified). If asked to pick a base, you want the current flagship generation, not v6-mini.
3. Upload your staged set — bulk, from your training folderUpload the exact files from QUELL-CORE-v1/. Minimum is 6 tracks; you are uploading 10–15. Upload takes a few minutes for a set this size. If the copyright screen flags a track, do not fight it in the moment — note it, swap in your prepared alternate, and see §11.
4. Name it like a version, not a vibeQUELL-CORE v1 — 2026-10, not “My Sound.” You will rebuild; the name is the first line of your model log. If the form offers a description or style notes, describe the set factually (era, lane, energy) — it is documentation for future-you.
5. Build and waitTraining typically completes in 2–5 minutes; reports range up to an hour under load or with large sets. Budget for the possibility of a creation cost in the v6 picker (100 credits has been reported Unverified — single source) — immaterial on Pro/Premier credit pools, but note what you were actually charged in your log.
6. Log the build immediatelyModel name, date, exact track list, file source folder, plan, and any flags or swaps. §09 has a copy-ready template. A model you cannot reconstruct is a model you cannot improve.
7. Do NOT judge it on the first generationOne generation is an anecdote. Go straight to the validation battery in §07 and generate the full test set before forming an opinion.

“Sounds kinda like me” is not a validation standard. Run all seven tests, in order, on the same day you build, while the training set is fresh in your mind. Rate each one as you go — the scorecard below keeps your verdict in this tab and gives you a ship / fix / kill call at the end. Generations spent here are the cheapest quality control you will ever buy: roughly 14–20 generations total. Expert practice

Everything you learned on 4.5+ still works — it just changes jobs. Your Style prompts used to carry two loads at once: what the artist sounds like and what this song should do. The Custom Model now owns the first load permanently. Prompting gets shorter, more specific to the song, and — this surprises people — more powerful, because every word you write is now about this track instead of re-establishing your identity from zero. Expert practice Community-tested

The new division of labor

Four habits that matter with a Custom Model active

• Stop describing your own sound back to it. If your prompt still contains your full identity stack (“gritty, analog-warm, wide choruses…”), you are double-teaching — and the prompt's version, being cruder, can fight the model's learned version. Brief the song, not the artist.
• Prompt the delta. The highest-value Style line is now the departure: “same world, but half-time feel and the sparsest verse arrangement yet.” The model supplies the gravity; you supply the vector.
• Use Exclude as a steering wheel, not a garnish. A strong home model will keep offering its favorite moves. When a generation repeats a signature element you are tired of hearing in this song, name it in Exclude (“no harmonica”, “no gang vocals”, “no half-time bridge”) rather than re-rolling and hoping. Precise negatives beat more adjectives.
• Match model variant to phase. Explore with v6-wild + your Custom Model — variation inside your identity, the most productive sketching mode v6 offers. Finish with v6 + your Custom Model and Variety pulled low: searching wants range, finishing wants repeatability. Use v6-mini only for throwaway structure sketches; do not judge the Custom Model by mini outputs. Expert practice

Then finish like a producer, not a gambler

v6's edge over 4.5+ is local control — use it after the Custom Model gets you close: plain-language section edits (“keep the verse, make the chorus wider, lift the vocal melody, same tempo”), single-lyric swaps without regenerating the track, and mashups built from your own catalog (your vocal idea over your unfinished arrangement is the honest, rights-clean use of the feature). The workflow inverts the old one: the model drafts in your voice, you edit like it's a session — instead of re-rolling a slot machine hoping it lands on you. Official

The constraint that shapes everything: once built, a Custom Model cannot be edited. There is no mechanism to add songs, remove songs, or reweight tracks after training. If the set needs changing, you build a new model. Slots are freed by deleting an existing model. Community-tested Treat each build like a release: versioned, logged, and reviewed — not a settings page you fiddle with.

Slot doctrine (3 slots, 3 jobs)

Rebuild discipline

• Change one variable per rebuild. Swap at most 3 tracks, targeting the specific fingerprint traits Test 3 found missing. Change the whole set and you learn nothing about what fixed (or broke) it.
• Name versions and never reuse a name: QUELL-CORE v1, QUELL-CORE v2 (swap: −Track A −Track B, +Track C +Track D). The name carries the diff.
• Re-run the full battery on every challenger — same prompts, same held-out witnesses. Scores are only comparable if the exam doesn't change.
• Promote or kill within a week. A challenger that sits unjudged becomes a third “which one did I like?” slot, and model confusion is how catalogs of generations turn to mush.
• Fold in new releases quarterly, not weekly. When a new finished track passes the §04 rubric at Identity 5, it becomes a rebuild candidate. Rebuilding for every new single churns slots and invalidates your comparisons.

Your situation — training on your own, already-released catalog — is the cleanest rights position this feature has. It is not a consequence-free one. Here is the honest map, in plain terms (not legal advice — for a catalog with real revenue, a music attorney reviews anything here you plan to rely on):

What you keep, and what you grant

• You retain ownership of what you upload. Suno's Terms (current version effective September 3, 2026) state users keep their rights in their Submissions; Suno does not claim ownership of them. Official
• But the license you grant is very broad: worldwide, perpetual, irrevocable, royalty-free, and sublicensable — covering use in connection with providing, monetizing, marketing, and improving Suno's products and AI models, with a moral-rights waiver where legally effective. This language is not new with v6; near-identical terms date to early 2024. Official
• Practical translation: you still own your masters and can exploit them anywhere, forever. What you cannot do is claw back Suno's permission to have used them — including as model-improvement data. A private Custom Model governs who can generate with your fine-tune (only you); it does not shrink the underlying license grant on the uploads themselves. Upload accordingly: released catalog = considered trade. Unreleased material you may shop, license exclusively, or keep in reserve = think twice.

Your released catalog, case by case

What comes out: your rights in generations

• Commercial rights attach to permitted downloads on a paid plan under the September 2026 terms — and persist for tracks downloaded while subscribed, even if you later cancel. Generation alone, or a free-tier trial download, carries no commercial rights. Mind the monthly download caps (Pro 20 / Premier 60) when planning a release batch. Official
• Commercial rights ≠ guaranteed copyright. Suno grants what rights it can contractually; whether an AI-assisted work is copyrightable, and how much human authorship it needs, is a jurisdiction question Suno explicitly does not warrant. Your substantial human input — your catalog as source, your lyrics, your editing and arrangement decisions — is both an artistic and a legal strength. Document it (your model log does double duty here).
• Your Custom Model itself is private and non-transferable — nobody else can generate with it, and it does not travel if you change accounts. If you publish a song that uses your Voice and allow remixing, others may be able to remix that song inside Suno — a separate permission to manage at publish time. Official
• v6's label partnerships do not clear your outputs. Warner/BMG/Believe involvement improves the platform's footing; it is not a blanket clearance for any individual track you generate, and other litigation (Universal, Sony) is still open. Release hygiene — original lyrics, no imitation briefs, distributor AI-disclosure where asked — stays your job. Official

One sitting, start to validated model. Everything before minute 55 is work you already know how to do — listening to your own records with a rubric instead of nostalgia.

1. Minutes 0–30 — Score the catalogSpreadsheet open, catalog playing on shuffle. Score Identity / Quality / Coherence for your strongest ~25 candidates. Mark 4 signature tracks as held-out witnesses. No Suno yet.
2. Minutes 30–45 — Pick and stage the setTop 12 by the rubric → copy the WAV masters into Drive: CustomModel/QUELL-CORE-v1/. Trim dead air, fix filenames, verify each file plays through.
3. Minutes 45–55 — Prep the exam before the student existsWrite the held-out brief (P2) and your five-trait fingerprint list. Copy the prompt set (§07) into a doc. Deciding what “good” means before you hear output is the whole discipline.
4. Minutes 55–65 — BuildPro/Premier confirmed → Custom Models → Create → upload the staged folder → name it QUELL-CORE v1 — 2026-10 → build. Log it with the §09 template while it trains.
5. Minutes 65–90 — Validate, don't vibeRun the prompt set on the model and on base v6. Score the seven tests in the §07 scorecard. Ship, challenger, or kill — by the decision rule, not by the best single generation you heard.

One last discipline note: this feature rewards the exact skill your catalog already proves — finishing records and making decisions about them. The artists who get the most from Custom Models won't be the best prompters. They'll be the best editors of their own history.