MiniMax Music 3 can finally run locally, hands-on guide to achieving AI music creation freedom
MiniMax Music 3 can finally run locally now — here's a hands-on guide to achieving true AI music creation freedom. When you open apps like Qishui Music or other listening platforms, you might notice that over 50% of the songs are now AI-generated, AI-arranged, or AI-remixed. The wave of AI music creation has arrived, and there's no stopping it now. So when I saw MiniMax M
💡 What You Will Learn
MiniMax Music 3 can finally run locally now — here's a hands-on guide to achieving true AI music creation freedom. When you open apps like Qishui Music or other listening platforms, you might notice t
📜 Table of Contents
MiniMax Music 3 Can Finally Run Locally — Here's My Hands-On Guide to AI Music Creation Freedom
If you've been using music streaming apps like 汽水音乐 lately, you've probably noticed that over 50% of the songs are now AI-generated, AI-remixed, or AI-covered. The wave of AI music creation is here, and there's no stopping it.
So when I saw the news that MiniMax Music 3 had released its weights, my first thought was: does this mean I can finally bring song generation back to my own machine and try it out?
What's the Big Deal About This Open-Source Release (Fully Local Deployment)
MiniMax Music 3 hit the API on July 16, but the model weights weren't made public until August 13. What does open weights mean? The entire song generation pipeline can now run locally — up to roughly 5 minutes of output, 32kHz stereo, complete with an official ComfyUI workflow.
It takes two inputs: lyrics and a music description. In the lyrics, you can use section tags like [Intro], [Verse], and [Chorus]. In the description, you can get as granular as vocal style, instrumentation, and arrangement direction. This feels a lot more like placing an order with a producer than telling a web app "give me a melancholic folk song."
Download 11.9GB, Squeeze It Into a 16GB GPU (Low-VRAM Setup)
The original repo shows roughly 57.4GB on the model page — clearly not sized for consumer GPUs. ComfyUI repackaged a low-VRAM bundle: the INT8 generation model (8-bit integer quantization, VRAM-friendly) is about 2.5GB, the text encoder about 9.2GB, and the audio decoder about 0.2GB — roughly 11.9GB in total. Combined with dynamic VRAM and CUDA Graphs (NVIDIA GPU acceleration), the goal is to get this running on a 16GB card.
How I Got It Running Locally (Step-by-Step)
Step 1: Download the three model files. Put the generation model in ComfyUI's diffusion_models directory, the text encoder in text_encoders, and the audio decoder in vae. My models live in a shared directory at C:/Users/yourname/ComfyUI-Shared/models — just drop them into your own ComfyUI model folders.
Step 2: Launch ComfyUI. In the terminal, I cleared PYTHONPATH first before starting the main program — otherwise it collides with another Python environment's libraries and won't boot. The command looks like env -u PYTHONPATH python main.py --listen 127.0.0.1 --port 8188 --disable-auto-launch. I also set up an extra_model_paths.yaml pointing to the shared model directory.
Step 3: Load the official workflow. You can find the MiniMax Music 3 Text to Music workflow in ComfyUI's template library, or import the official JSON directly.
Step 4: Fill in two inputs. The Caption field takes the music description — the official recommendation is to split it into three parts: global music info (tempo, key, emotional arc), vocal details (timbre, singing style, harmonies), and arrangement (instruments, section changes). The Lyrics field takes your lyrics, with section tags like [Intro], [Verse], and [Chorus] at the start of each section.
Step 5: Configure parameters. For low VRAM, switch the generation model to the INT8 version, enable tiled decode (block-wise decoding to save VRAM), set the melody duration to 60 seconds initially, and fix a Seed value for reproducibility.
Step 6: Submit, then wait. The task took 150.05 seconds from submission to completion, outputting an MP3 of about 1.9MB.
My Real-World Test (RTX 4060 Ti 16GB)
My rig: 32GB RAM + RTX 4060 Ti 16GB, ComfyUI 0.31.0, PyTorch 2.12.1. I wrote some Chinese lyrics — the theme felt a bit like my own state at that moment: late night, screens, command lines. Modest compute, but enough. I set 60 seconds, 30 sampling steps, and a fixed seed. The task took 150.05 seconds from submission to completion, producing an MP3 of about 1.9MB. Generation time was roughly 2.5x the target duration.
That "up to 5 minutes" note on the page doesn't exactly fill me with confidence. 60 seconds already takes 150 — stretch it to 5 minutes and you'll need to re-validate time, memory, and audio quality. Just running that 60-second clip had the GPU near maxed out, but the whole task never hit an out-of-VRAM error, so I'm honestly quite happy with the result.
What Does Local Deployment Cost You?
If you just want to generate a song from a web page, Music 3's API has been available for a while. Local deployment is for a different crowd: people who want full control over generation parameters, want to do secondary development, don't want to hand lyrics and audio to the cloud, or care deeply about IP and privacy.
But the price is having a decently specced machine, a solid GPU, and being willing to manage model directories, ports, and error messages yourself. Just this one session, I hit three snags — environment variables, workflow format, and required parameters — all of which I had to figure out on my own. This is a completely different beast from clicking a few buttons on a website. It does take some effort.
My Take (Who This Is For)
I'm on board with this direction — MiniMax has handed full song generation capability over to local machines. For anyone into AI music creation, this is a godsend: unlimited attempts, no pressure from credit consumption, and your only creative cost is the electricity bill. For beginners in AI creation, it also removes the cost barrier entirely.
That said, the system requirements are steep and there are plenty of pitfalls — but at least the pipeline can be made to work with some effort. If you're comfortable tinkering locally and willing to invest time in AI music creation, MiniMax Music 3 deployment is well worth the hassle.
In 2026, would you go through the trouble of setting up a local model just to play with AI music creation? Let's chat in the comments. If you need the ComfyUI workflow, feel free to DM me.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
