MiniMax Music 3 on 8GB VRAM: Full Songs With Vocals
Want to go deeper than this article?
Free account unlocks the first chapter of all 25 courses — RAG, agents, MCP, voice AI, MLOps, real GitHub repos.
Got the hardware sorted? Now build on it. You know what to buy — the courses show you what to actually run, fine-tune, and ship on it. First chapter free, no card.
Short answer: yes, MiniMax Music 3 runs on an 8GB card — but the 8GB path is a streaming trick, not a fit. MiniMax's own model card says full precision "fits under 24GB", drops to ~22GB with CPU offload, and only reaches 8GB by streaming the language model layer by layer. The thing eating your VRAM is not the diffusion model (2.5GB at int8) — it is the 18.47GB text encoder. Output is 32kHz 16-bit stereo, up to five minutes, with sung vocals.
That is the whole page in a paragraph. What follows is the file-by-file detail: exactly which weights to download for your card, where the ComfyUI route differs from the diffusers route, and the two clauses in the licence you need to read before you put a single track on Spotify.
One disclosure up front, because it is load-bearing: we have not yet timed generation on an 8GB card ourselves. The 8GB claim on this page is MiniMax's, quoted from the model card. Every file size, licence term, and architecture number below is read straight off the official repositories. Where a number is a vendor claim, it is labelled as one.
What It Is
MiniMax Music 3 is an open-weights text-to-music model that writes and sings complete songs — lyrics, vocals, arrangement — from two text inputs. It landed on Hugging Face on 7 August 2026 and has been updated since (last revision 14 August). Reading the Hugging Face counters on 18 August 2026, the main repo showed roughly 10,400 downloads and 920 likes, with a community GGUF conversion at about 2,700 downloads. Those are 30-day rolling figures and they move daily — treat them as a snapshot, not a scoreboard. For a music model two weeks old, it is fast uptake.
The architecture, per the model card:
| Component | Size | Job |
|---|---|---|
| Global LLM | 8B (initialised from Qwen3-8B) | Long-range structure; predicts the first RVQ codebook frame by frame |
| Local LLM | 0.6B | Frame-level acoustic detail; the remaining codebooks |
| Flow Matching | 2.4B | Continuous hidden-state synthesis |
| Flow-VAE decoder | 123M | Final 32kHz stereo audio |
The tokeniser uses eight layers of residual vector quantisation — a 16,384-entry semantic codebook plus seven 1,024-entry acoustic codebooks. At inference the synthesis path fuses the LLM hidden states directly and skips the discrete tokeniser decoder entirely.
What that buys you in practice is the thing every previous open music model failed at: a song that still remembers its own chorus four minutes in. Intro, verse, pre-chorus, chorus, bridge, instrumental, outro — with a consistent vocal identity across the whole thing.
Reading articles is good. Building is better.
Free account = the first chapter of all 25 courses, with a per-chapter AI tutor. No card.
Will It Fit? Per-VRAM Table
Pick your row, download those three files, ignore everything else. These sizes are read directly from the Comfy-Org repack and the community GGUF repo file listings — they are exact, not estimates.
| Your GPU | Diffusion model | Text encoder | VAE | On-disk | Reality check |
|---|---|---|---|---|---|
| 8GB (3060 Ti, 4060, 3070) | GGUF Q4_K_M — 1.49GB | pruned int8 — 9.20GB | 0.217GB | ~10.9GB | Works only with aggressive offload; the encoder lives in system RAM. Slow. |
| 12GB (3060 12GB, 4070, 5070) | int8_convrot — 2.50GB | pruned int8 — 9.20GB | 0.217GB | ~11.9GB | The sensible floor. Still offloading, but far less thrash. |
| 16GB (4060 Ti 16GB, 4080, 5070 Ti) | fp16 — 4.91GB | pruned int8 — 9.20GB | 0.217GB | ~14.3GB | First tier where the whole stack is nearly resident. |
| 24GB (3090, 4090, 5090 has more) | fp16 — 4.91GB | pruned bf16 — 16.71GB | 0.217GB | ~21.8GB | Matches the card's "full precision fits under 24GB". |
| 32GB+ (5090, RTX Pro, A6000) | fp32 — 9.83GB | bf16 — 18.47GB | 0.217GB | ~28.5GB | No compromises, no offload, longest songs. |
| Apple Silicon | — | — | — | — | No supported path. The card lists "Inference requires CUDA". |
The community GGUF ladder for the diffusion transformer, if you want to trade fidelity for headroom (sizes from the repo listing):
| Quant | Size | Repo's own note |
|---|---|---|
| F16 | 4.98GB | Uncompressed reference baseline |
| Q8_0 | 2.70GB | Near-lossless (the repo's recommended pick) |
| Q6_K | 2.12GB | Excellent fidelity |
| Q5_K_M | 1.79GB | Balance of speed and footprint |
| Q4_K_M | 1.49GB | Flagged for 8GB–10GB GPUs |
| Q3_K_M | 1.16GB | "May introduce audible softening" |
System RAM matters more here than on any LLM you have run. Everything that does not fit in VRAM sits in RAM and gets streamed across PCIe every step. Our working rule for the 8GB and 12GB rows: budget 32GB of system RAM, and do not attempt it on 16GB while a browser is open. That is our guidance from how offloading behaves generally, not a number from MiniMax.
If your card is the constraint rather than your patience, the hardware hub and our GPU ranking for local AI cover what a step up actually buys. A used 24GB card — see the RTX 3090 guide — puts you straight into the "no compromises" row for less than a new mid-tier GPU.
The Text Encoder Trap
The single most useful thing to understand: MiniMax Music 3's "text encoder" is an 8B language model, and it is 80% of your memory problem.
Look at the numbers side by side. Comfy-Org ships three text encoders:
| Text encoder file | Size |
|---|---|
minimax_music3_text_encoder_bf16.safetensors | 18.47GB |
minimax_music3_text_encoder_pruned_bf16.safetensors | 16.71GB |
minimax_music3_text_encoder_pruned_int8_convrot.safetensors | 9.20GB |
And three diffusion models:
| Diffusion model file | Size |
|---|---|
minimax_music3_dit_fp32.safetensors | 9.83GB |
minimax_music3_dit_fp16.safetensors | 4.91GB |
minimax_music3_dit_int8_convrot.safetensors | 2.50GB |
The 18.47GB "text encoder" lines up almost exactly with the 18.48GB qwen_7B directory in the original repo. It is the Qwen3-8B-derived Global LLM wearing a ComfyUI hat.
This is why the GGUF quants, useful as they are, do not solve the 8GB problem on their own. Going from F16 (4.98GB) to Q3_K_M (1.16GB) saves you 3.8GB. Going from the bf16 encoder to the pruned int8 encoder saves you 9.3GB. Get the encoder choice right first; the DiT quant is a rounding error next to it.
Practical consequence for the 8GB row: even the smallest encoder is 9.20GB, which is bigger than your entire card. Nothing you do makes it resident. The model runs because ComfyUI and diffusers page pieces of it in and out — which is exactly what MiniMax means by "streaming the language model layer by layer".
Setup: ComfyUI
ComfyUI has native support and an official template — this is the easiest route, and the only one with a low-VRAM story that does not involve writing Python.
Per the ComfyUI docs, the flow is: update ComfyUI to the latest version, open Template Library → Audio → MiniMax Music 3, and let the pop-up fetch the models. If the nodes are missing, your ComfyUI is too old.
If you would rather pull the weights yourself, these are the exact commands from the repos:
# Diffusion model (int8 — the 12GB pick)
huggingface-cli download Comfy-Org/MiniMax-Music-3 \
diffusion_models/minimax_music3_dit_int8_convrot.safetensors \
--local-dir ComfyUI/models/diffusion_models
# Text encoder (pruned int8 — the low-VRAM pick)
huggingface-cli download Comfy-Org/MiniMax-Music-3 \
text_encoders/minimax_music3_text_encoder_pruned_int8_convrot.safetensors \
--local-dir ComfyUI/models/text_encoders
# VAE
huggingface-cli download Comfy-Org/MiniMax-Music-3 \
vae/minimax_music3_dav.safetensors \
--local-dir ComfyUI/models/vae
For the GGUF route you also need the ComfyUI-GGUF custom node, and you load the model through Unet Loader (GGUF) instead of the standard loader:
huggingface-cli download Abiray/MiniMax-Music3-GGUF \
MiniMax-Music3-Q4_K_M.gguf \
--local-dir ComfyUI/models/diffusion_models
Two workflow controls are worth knowing before your first run, both documented by ComfyUI:
max_duration— target length in seconds. The template defaults to 60; the model supports up to about 300. Longer costs more time and more VRAM.tiled_decode— decodes the audio VAE in overlapping tiles to cut VRAM. ComfyUI's own note: helpful for long songs on low-VRAM GPUs, slightly slower, small risk of seams at tile boundaries. Turn it on for the 8GB and 12GB rows. Turn it off above that.
New to ComfyUI generally? Start with our ComfyUI complete guide. If it loads and then dies mid-generation, that is almost always a memory-management problem rather than a MiniMax problem — our ComfyUI out-of-memory guide covers the usual causes.
Run this on your own machine and stop paying every month
Pay once and keep it. No renewal, no per-token bill, and nothing you feed it ever leaves your hardware.
Setup: Diffusers and SGLang
If you want the 8GB path exactly as MiniMax describes it, you need the diffusers route — the group-offloading hook is what makes the claim true.
At the time of writing the diffusers integration is still an unmerged PR, so the card pins a specific commit:
pip install git+https://github.com/huggingface/diffusers@dafe3733fcfdbf3c48915fe77be3aef65b5d6a2d \
transformers accelerate soundfile
The card's own low-VRAM snippet, which is the load-bearing part:
import torch
from diffusers import ComponentsManager, ModularPipeline
from diffusers.hooks import apply_group_offloading
manager = ComponentsManager()
manager.enable_auto_cpu_offload(device="cuda")
pipe = ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-Music3", components_manager=manager)
pipe.load_components(dtype=torch.bfloat16)
# Only needed below ~22 GB of VRAM — slower, but fits in 8 GB.
apply_group_offloading(
pipe.language_model, onload_device=torch.device("cuda"),
offload_type="leaf_level", use_stream=True
)
Note the comment MiniMax left in their own example: "slower, but fits in 8 GB." They are not overselling it.
There is also an SGLang-Omni server path, which is the one to use if you are generating in volume rather than experimenting:
hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm
sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000
It exposes an OpenAI-shaped speech endpoint — lyrics go in input, the music description goes in instructions:
curl http://127.0.0.1:8000/v1/audio/speech \
-H 'Content-Type: application/json' \
-d '{
"model": "MiniMaxAI/MiniMax-Music3",
"input": "[Verse]\nMorning light filtering through the pine\n[Chorus]\nSoftly the world begins to breathe",
"instructions": "A warm acoustic pop song with intimate female vocals, fingerpicked guitar, soft piano, and a gradual emotional build into a wide final chorus.",
"response_format": "wav",
"seed": 7,
"max_new_tokens": 750,
"stream": false
}' --output minimax_music3.wav
max_new_tokens counts audio frames at 25 frames per second, so 750 frames is 30 seconds. A five-minute song is 7,500 — and the hard ceiling is 9,000 frames.
Be aware of the download. hf download on the original repo pulls roughly 57GB (we summed the file listing: 18.5GB Qwen encoder, 17.2GB language model, 9.8GB Flow-VAE, 9.7GB transformer, 1.3GB RVQ depth decoder, plus the rest). The ComfyUI repack exists precisely so you do not have to do that.
Prompting: Caption + Lyrics
Two inputs, and the model is far more responsive to the caption than most people expect. MiniMax recommends a structured caption in three named sections:
- Global Metadata — genre, subgenre, BPM, key, scale, emotional progression, listening scenario, production profile
- Vocal Details — vocal gender, timbre, performance style, harmony, backing vocals, vocal effects
- Arrangement — primary and secondary instruments, section-level instrument evolution, groove, bass, percussion, textures, spatial effects
Lyrics carry the structure through section tags on their own lines: [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], [Outro].
The example from the official repo, verbatim:
Genre: acoustic pop. BPM: 96. Key: C major. Warm and intimate, building gently
into the chorus. Vocals: soft female lead, close and breathy, light stacked
harmonies in the chorus. Arrangement: fingerpicked guitar and soft piano;
brushed drums and upright bass enter in the chorus.
MiniMax also ship a music-caption-rewriter skill that expands a one-line idea into the full structured caption:
npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter
One honest caveat from the model card's own limitations list: section tags and descriptions provide generative control rather than strict symbolic guarantees. Ask for 96 BPM in C major and you will usually get something near it, not a metronome-accurate render. Plan on re-rolling seeds.
The Licence: Read This Before You Monetise
MiniMax Music 3 is not open source in the OSI sense. It ships under the MiniMax-Music3 Community Licence, and commercial use carries two specific obligations. Quoting the LICENSE file directly:
- Attribution is mandatory and visible. "You shall prominently display 'MiniMax-Music3' on the user interface of commercial product or service that uses the Software." Not a footnote in your terms — on the interface.
- A $20M revenue ceiling. Above "20 million US dollars" in aggregate yearly revenue across you and your affiliates, you need "separate, prior written authorization from MiniMax" via api@minimax.io.
- Safeguards if you host it for others. If you offer a product or service that lets third parties generate outputs, you must implement, maintain, test and periodically review "reasonable and proportionate technical and organizational safeguards" against infringing outputs — and not weaken them.
- An Acceptable Use Policy MiniMax can revise. Exhibit A runs to 19 clauses. Two that bite ordinary creators: no undisclosed machine-generated content in public environments, and no impersonation without consent.
For most independent producers, clauses 1 and 4 are the whole story, and clause 2 will never apply. That is a genuinely usable deal — considerably better than the research-only licences elsewhere in this space. But it is a licence with conditions, not a permissive one, and "prominently display MiniMax-Music3" is a real product-design constraint if you were planning to white-label a song generator.
One correction worth flagging. The community GGUF repository tags itself license: apache-2.0 and its README says it "inherits the Apache-2.0 License from the original release". That is incorrect. The original release carries the MiniMax Community Licence — the LICENSE file is right there in the repo. Do not rely on a third-party quantiser's frontmatter for your legal position. (Underlying components have their own permissive licences — Qwen3-8B is Apache-2.0, the DiT derives from Stable Audio under MIT, the VAE from DAC under MIT — but the composed model is governed by MiniMax's terms.)
If licence terms are the deciding factor for you, the contrast with the voice side of this stack is stark: VoxCPM2 is straight Apache-2.0, no attribution clause, no revenue ceiling.
Honest Limitations
Straight from the model card's own limitations list, plus what follows from the file sizes:
- CUDA required. No Apple Silicon, no ROCm path documented. This is the biggest exclusion.
- Non-streaming generation only. You wait for the whole song; there is no play-as-it-renders.
- 5,000-token prompt cap, and a 9,000 acoustic frame ceiling (360 seconds).
- Control is soft. Tempo, key, instrumentation, lyrics and structure "may not always match every requested detail exactly".
- The 8GB path is a fallback, not a target. MiniMax's own code comment calls it slower. We have not measured how much slower.
- Nothing here is a mixing engineer. You get a finished-sounding stereo file, not stems.
What we could not verify and are not going to guess at: generation time per minute of audio on any specific card, real peak VRAM under load (as opposed to on-disk file sizes), and quality deltas between the GGUF quant levels. When we have run it on our own hardware, this page gets the numbers.
vs ACE-Step and Suno
Aimed squarely at long-form coherence and vocals, where the open incumbents are weakest — but ACE-Step is still the far bigger download, and nobody has published a blind listening test either way. Against Suno you give up polish and convenience; you get weights you keep.
Against ACE-Step, the incumbent open text-to-music model: be clear that ACE-Step is still the far bigger download. On 18 August 2026 the Hugging Face 30-day counters put ACE-Step 1.5 at roughly 57,000 on the main repo, with its ComfyUI repack and community GGUFs adding several hundred thousand more, against MiniMax Music 3's ~10,400. MiniMax is the newcomer, not the incumbent. What it is aimed at is exactly ACE-Step's weak spot — five-minute structural coherence with a stable vocal identity, rather than a good-sounding 30-second loop. The trade is size: ACE-Step runs comfortably where MiniMax Music 3 needs offloading. We have not run a blind listening comparison between the two, and we are not going to imply one.
Against Suno: Suno wins on polish, iteration speed, and not requiring a GPU. MiniMax Music 3 wins on the three things that make people search for a local alternative in the first place — no per-song credit meter, no upload of your lyrics to someone's server, and weights you keep. If you generate five songs a month, Suno is cheaper and better. If you generate five hundred, or your lyrics are under NDA, or you want to fine-tune, this is one of the few open models where that trade is genuinely close.
For the broader landscape of local music tooling, see our local AI music and DJ tools roundup and the music generation tutorial — both predate this model, and MiniMax Music 3 is now the answer they were waiting for.
Verdict
- The 8GB headline is true but load-bearing on a caveat. It fits by streaming an 8B language model layer-by-layer from system RAM. It will finish. We cannot yet tell you how long you will wait.
- 12GB is the honest floor, 16GB is comfortable, 24GB matches the card's "full precision" tier. Pair any of them with 32GB of system RAM.
- Choose the text encoder first. Pruned int8 (9.20GB) versus bf16 (18.47GB) is a 9GB decision. The DiT quant is worth 2-4GB.
- Use ComfyUI unless you have a reason not to. Native template, tiled decode, ~12GB of downloads instead of 57GB.
- Read the LICENSE, not the GGUF repo's frontmatter. Community licence, visible attribution required, $20M revenue gate. Usable for almost everyone — but it is a condition, not a freedom.
"Make a full song with vocals, on my own machine, and sell it" is a real sentence now — though be honest that MiniMax Music 3 is not the only model it applies to. ACE-Step 1.5 carries a plain MIT licence, which is strictly less restrictive than MiniMax's community terms. What MiniMax brings is the architecture built for five-minute structure with one consistent voice. The hardware bill is steeper than the headline suggests, and the licence has strings. Both are still a much better deal than a credit meter.
FAQ
Got the hardware sorted? Now build on it.
You know what to buy — the courses show you what to actually run, fine-tune, and ship on it. First chapter free, no card.
Decide before you spend a thousand pounds
The AI Hardware course sizes your build properly — VRAM ladder, real bottlenecks, budget builds — and Pick the Right Model tells you what to run on it.
Liked this? 25 full AI courses are waiting.
From fundamentals to RAG, agents, MCP servers, voice AI, and production deployment with real GitHub repos. First chapter free, every course.
Build Real AI on Your Machine
RAG, agents, NLP, vision, and MLOps - chapters across 25 courses that take you from reading about AI to building AI.
Want structured AI education?
25 courses, 519+ chapters, from $9. Understand AI, don't just use it.
Continue Your Local AI Journey
- PILLARLocal AI Hardware Requirements (2026): Complete Guide
- AI Hardware Guide 2026: GPU, CPU & RAM for Local AI
- AI Hardware Requirements: CPU, GPU and RAM for Beginners
- AI RAM Requirements 2026: How Much for 7B, 13B, 70B Models?
- AI Server Build Under $1,500: Parts List and What Fits
- AMD Ryzen AI Max+ 395 (Strix Halo) for Local AI 2026
- Apple M4 for Local AI: Mac Studio + MacBook Guide (2026)
- Benchmark Your Local AI Setup: tok/s, TTFT, VRAM
- Best GPU for AI Video Generation: By VRAM Tier (2026)
- Best Local AI Models 2025: 6 Compared (RAM, VRAM, MMLU)
Comments (0)
No comments yet. Be the first to share your thoughts!