MiniMax-Music3 Requirements: VRAM, License & API Sunset

MiniMax AI created a Hugging Face repository named "MiniMaxAI/MiniMax-Music3" on August 7, 2026, and published the weights with a last update on August 14, 2026 (the main branch totals 57.4GB). This is an open-weight text-to-audio model: given a structured caption plus lyrics annotated with section tags, it outputs a full track up to 5 minutes long as 32kHz/16-bit stereo WAV. Around the same time, MiniMax's official documentation announced that the paid music-generation API will stop accepting new users starting August 20, 2026. This article covers both sides — the confirmed local hardware requirements for running the open weights yourself, and the API situation — strictly within what is publicly documented. If you're evaluating running large generative models locally, it's worth comparing this against the VRAM requirements for the video model MiniMax H3 (../columns/minimax-h3-requirements-vram-local-2026) and LTX-2.5 (../columns/ltx-2-5-requirements-vram-local-2026).

Per the README's own summary, MiniMax-Music3 is a hierarchical autoregressive system combining an "8B Global LLM (long-range song structure)" with a "0.6B Local LLM (frame-level acoustic detail)", plus Flow Matching and a Flow-VAE for synthesizing continuous hidden states. Specifically, the Global LLM (8B) predicts the first RVQ codebook for each frame, while the Local LLM (0.6B) predicts the remaining acoustic codebooks within each frame. Flow Matching is sized at 2.4B and the Flow-VAE Decoder at 123M. Notably, the Hugging Face model card metadata displays "2B params", which does not match the README's own component breakdown (8B+0.6B+2.4B+123M). Public information does not clarify which figure is the mislabel; this article records the discrepancy as a fact without speculating on its cause.

The README does not include any speed benchmarks (how many seconds it takes to generate a track of a given length), so per-GPU generation times are not publicly confirmed. It's reasonable to assume lower-VRAM setups run slower, but that is a general inference, not a measured figure for this specific model. If you're comparing local inference stacks more broadly, see the local LLM inference engine comparison (../columns/local-llm-inference-engine-comparison-2026).

MiniMax-Music3 takes two kinds of input. The first is a structured music caption with Global Metadata, Vocal Details, and Arrangement sections. The second is lyrics, annotated with section tags such as [Intro], [Verse], [Chorus], [Bridge], [Instrumental], and [Outro]. An example of tagged lyrics is shown below.

The official Comfy-Org/MiniMax-Music-3 tutorial distributes several weight variants: DiT in fp16/fp32/int8_convrot (for low VRAM), the text encoder in bf16/pruned bf16/pruned int8_convrot, and the VAE (dav). The tiled_decode option can substantially reduce VRAM usage, but the docs explicitly note it may introduce slight seams at tile boundaries. The workflow template defaults to 60 seconds, with a maximum of roughly 5 minutes (300 seconds), and "longer tracks consume both more time and more VRAM" per the documentation. Output is saved as MP3 files under ComfyUI/output/audio/. The Comfy-Org README does not list GB sizes for the individual weight files.

MiniMax-Music3 ships under the "MiniMax-Music3 Community License", a custom license rather than Apache 2.0 or MIT. Commercial use is allowed, but the license requires prominently displaying "MiniMax-Music3" on the user interface of any commercial product or service that uses it (verbatim: "You shall prominently display 'MiniMax-Music3' on the user interface of commercial product or service that uses the Software."). If your annual revenue exceeds US$20 million, you must contact api@minimax.io with the subject "MiniMax-Music3 licensing - authorization request" and obtain prior written authorization. The license also includes an obligation to implement misuse safeguards and an Acceptable Use Policy covering 19 prohibited-use categories. No regional restriction (such as excluding the US or EU) was found in the LICENSE text.

MiniMax's official documentation states: "Starting August 20, 2026, the paid APIs (Music Generation and Lyrics Generation) will no longer be available to new users; existing paying users can continue to use the current API services. The free music generation APIs (Music-3.0-free, Music-2.6-free, music-cover-free) will be discontinued." In short, from August 20, 2026 onward, the paid music- and lyrics-generation APIs stop accepting new users (existing paying users can keep using them), and the free tiers (Music-3.0-free, Music-2.6-free, music-cover-free) are discontinued entirely. For anyone starting fresh after that date, the practical choices narrow to either using MiniMax Audio (the vendor's own service) or running the open weights locally. Third-party media outlets have reported a price of "$0.15 per generation (up to 5 minutes)", but this figure does not appear on the relevant official documentation page and should be treated as unverified third-party information.

- An official list of supported languages (the README has no explicit enumeration; reports of Japanese vocal support exist but are not confirmed by primary sources)
- Concrete generation speed benchmarks (seconds per GPU tier)
- The breakdown or provenance of the training data
- Why the HF metadata's "2B params" figure diverges from the README's own component breakdown (8B+0.6B+2.4B+123M)
- Official confirmation of the third-party-reported "$0.15 per generation" price