MiniMax Music 3 vs. ACE-Step 1.5 XL

Two philosophies of making a song

These models disagree about what music generation is. ACE-Step 1.5 XL treats a song as one big canvas: a 4B language model sketches the semantic plan, then a 4B diffusion transformer denoises the entire track in 50 steps, whole-song-at-once. MiniMax Music 3 treats a song as a story told in order: an 8B language model (a retuned Qwen3-8B) writes the track frame-by-frame, 25 frames per second of audio, each frame refined by a small depth decoder across 8 RVQ codebooks — and only then does a 2.4B flow-matching transformer polish the result into the final waveform.

The test: six briefs, one GPU, official recipes only

We rendered six matched briefs on both models — identical lyrics, identical creative intent, translated into each model's native prompt idiom (structured caption for Music 3, tag list plus BPM/key for ACE). Anthemic pop-rock with a female lead, lo-fi instrumental, cinematic orchestral, a male-crooner jazz ballad, festival house, and a boom-bap rap verse. Full-length takes this time — every test on this page runs at least two minutes — with fixed seeds, rendered back-to-back on the same RTX 5090 through the same Sogni worker image, saved straight to MP3.

Both models ran their official recipes, unmodified: ACE-Step 1.5 XL Turbo — the stable production build, not the experimental SFT — at 8 steps / CFG 1 / shift 3 / euler-simple, and Music 3 at 30 steps / CFG 1.7 / euler-simple with planner CFG 1.7 and top-k 50 — the exact settings in each model's official ComfyUI template, on the exact int8 planner variant production serves. No cherry-picking: every take is the first render of its brief; the only re-roll rule is length (anything under two minutes gets a larger ceiling and a fresh take, noted where it happened).

One structural difference shows up before you press play: ACE fills exactly the duration you ask for, every time — 2:30 on the nose, all six briefs. Music 3 treats the requested duration as a ceiling and writes its own ending, so its takes land wherever the arrangement resolves.

Speed: the composer makes you wait for it

Autoregression has a price. Music 3's planner emits 25 frames for every second of audio — a two-and-a-half-minute song is 3,750 sequential passes through an 8B model before diffusion even starts. ACE-Step XL Turbo denoises the entire song in eight parallel steps.

One integration lesson hides in that table, and it nearly fooled us: our first int8 measurements were 4× worse than the numbers above, because the quant kernels silently ran in eager PyTorch mode on a CUDA-12 PyTorch build — comfy-kitchen only enables its optimized CUDA backend on cu130+. If your int8 renders seem inexplicably slow, read your startup log: a backend cuda: disabled line is the whole story. Benchmark the stack you'll actually ship, then benchmark it again on the stack you thought you were shipping.

Listening: six briefs, twelve takes

Every player below is wired to a real first-take render on the recipes above. We're not going to tell you what to hear before you press play — each pair notes what the brief stresses, and the verdict waits until the end.

The range: twelve more songs, one model

A versus test tells you who wins a fight; it doesn't tell you how big the winner's world is. So we pointed Music 3 at twelve briefs chosen to stress everything the six-genre suite doesn't: vocals in three languages, a gospel choir in call-and-response, a male–female country duet, an a cappella group with beatboxed percussion, metal, slow blues, two instrumental electronic idioms, a genre fusion that shouldn't work, solo piano with nothing to hide behind — and one full three-minute song with an intro, two verses, a bridge, and an outro. Every take below is the first render of its brief on the production int8 recipe, fixed seed, no cherry-picking; the only re-roll rule is length (every test on this page must clear two minutes). Durations still vary because Music 3 ends a song when the arrangement resolves — we asked for ceilings, it wrote endings.

What the community is hearing

We are not the only ones running this comparison. Across multiple independent Reddit threads and community reviews, a consistent picture has emerged — and it largely matches what our matched briefs surface.

On sound quality, Music 3 wins. The recurring praise: cleaner instrument separation, more realistic mixing and mastering, better-sounding drums and bass, and higher perceived production quality — especially for pop and orchestral material. "The instruments sound way better." "Sounds better than AceStep XL." "Best open source music generator." Community consensus puts its strongest genres at pop, cinematic, orchestral, piano, classical, and Chinese music — which tracks with our Mandarin ballad and trailer-cue takes above.

ACE-Step keeps real advantages. Users consistently credit it with more creativity, more unusual song structures, broader language coverage, finer prompt control, and editing workflows Music 3 simply doesn't have — cover, repaint, and section-level rework. One commenter put it neatly: "AceStep is ahead in creativity, MiniMax has better sound quality."

Vocals split the room. Many hear Music 3's vocals as smoother, less synthetic, and more cleanly produced; others find them emotionally flatter and too locked to the rhythm grid — "composition is solid but dynamic emotion is lacking," as one reviewer wrote. Our belt-vocal pop-rock pair above is a good place to form your own opinion.

Reported weak spots: metal, rock, EDM, and experimental genres. We deliberately kept a melodic-metal brief and a festival-house brief on this page so you can judge the community's least-favorite territory yourself rather than take anyone's word for it.

On speed, community reports for consumer GPUs put equivalent-length songs at roughly 30–60 seconds on ACE-Step 1.5 XL versus 120–180 seconds on Music 3 — a 3–5× gap. Against XL Turbo's eight-step recipe specifically, our measured gap is wider (see the table above); against ACE's higher-step configs it narrows toward the community's number. Either way the direction is the same: Music 3 charges you real GPU time for its fidelity.

Control surfaces: musician knobs vs. producer prose

ACE speaks musician: you hand it a BPM, a key, a time signature, a lyric language, and it obeys — which makes it the better instrument when a video already has a tempo grid or a brand spot needs exactly 30.0 seconds. Music 3 speaks producer: you describe the record — "female belt, doubled hooks, stripped-back bridge, tape saturation" — and it interprets. It also reserves the right to end the song early when the arrangement resolves, which is musically honest and occasionally inconvenient.

One structural note: Music 3 caps at 5 minutes; ACE goes to 10. And because Music 3's planner is literally a language model, lyric adherence and section structure are where it should earn its render time — that's exactly what the listening section above tests.

What this costs on Sogni

Pricing follows the physics. A 60-second Music 3 render costs about 51 Spark (~$0.26); the same minute on ACE-Step 1.5 XL Turbo is about 2.2 Spark (~$0.011). That's roughly a 23× premium — Turbo denoises a whole song in eight parallel steps while Music 3's 11B composer writes it frame by frame, and the meter reflects it. Both scale linearly with duration (and with extra steps, if you raise them): a two-and-a-half-minute Music 3 song runs about 128 Spark (~$0.64) versus ~5.4 Spark (~$0.03) on Turbo, and a full five-minute Music 3 song tops out around 256 Spark (~$1.28).

One quoting subtlety worth knowing: you're billed on the duration you request, and Music 3 treats that as a ceiling — it may resolve the song early. Ask for the length you actually want rather than defaulting to the maximum.

Notes for prompt writers

Music 3 wants a structured caption, not a tag soup. Write it in three movements, and label them: Global Metadata (genre, BPM, key, emotional arc), Vocal Details (voice type, delivery, diction — or “none, purely instrumental”), and Arrangement (instruments and how sections evolve). Every brief on this page uses exactly that shape.

Lyrics get their own field, with section tags on their own lines — [verse], [pre-chorus], [chorus], [bridge], [outro], or [instrumental] for no vocals. The planner genuinely reads them: it places your words in the sections you tagged, and the three-minute cut above followed an eight-section sheet through to its outro. Vocals follow the language of the lyrics you write — English, Mandarin, Spanish, and Japanese all worked first-take above.

A length lesson we learned the honest way: with a bare [instrumental] lyric field, the composer resolves instrumental pieces early — our lo-fi and solo-piano briefs came back at twenty seconds from a three-minute ceiling. The fix is structure: give instrumentals a skeleton of empty section tags ([intro] [verse] [chorus] [solo] [bridge] [outro]) and open the caption with the duration ("a full three-minute piece…"), and they play out their whole arc. Vocal tracks pace themselves by the lyric sheet — write enough verses for the length you want.

Three behaviors to plan around. The duration you set is a maximum — the composer writes an ending and may stop early, so describing a structure that fills your target length matters more than the number itself. One seed drives both the planner and the diffusion sampler, so a seed genuinely reproduces a take. And the official recipe (30 steps, euler/simple, CFG 1.7, planner CFG 1.7, top-k 50) is what we ship as defaults — raising top-k buys variety, not quality.

Verdict

MiniMax Music 3 is the new quality bar for open-weight music on the Sogni Supernet. Across the six matched briefs and the twelve-genre showcase it delivered cleaner mixes, more believable vocals in four languages, and song structures that genuinely follow the lyric sheet — all eight sections of the three-minute cut, through to the outro. That is exactly what an 11B autoregressive composer should buy with all those sequential passes, and it is why Music 3 launches as Sogni's premium music tier.

On sound quality and production quality, Music 3 does dethrone the workhorse. In these first-take renders it delivers cleaner mixes, more believable vocals, stronger separation, and more coherent long-form song structure. ACE-Step 1.5 XL Turbo remains exceptionally competitive — and it is the better, more efficient choice for many workflows. It renders the same minute of music at roughly 1/23rd the price and more than ten times the speed, hits an exact requested duration every time, and keeps the musician-grade controls — BPM, key, time signature — and editing workflows Music 3 simply doesn't have. Reach for Music 3 when the song is the product and final-record fidelity matters most. Reach for ACE-Step when speed, exact duration, control, editing, or drafts by the dozen matter more. The Supernet now serves both, and every take above stays up so you can re-run the comparison by ear.

Sources and further reading

• MiniMax-Music3 model card (MiniMax AI)
• Comfy-Org repackaged weights
• ComfyUI MiniMax Music 3 tutorial
• Comfy Blog: MiniMax Music 3
• Official Music 3 demo page
• ACE-Step on Hugging Face

More from Sogni

• Your Prompt Is Now a Director — MiniMax H3 on Sogni
• Upscale to 8K Without Changing Your Image — RTX Video Super Resolution on Sogni
• Powering an Unlimited AI Economy That Works