이미지: METAL LAB 생성
Summary
- MiniMax released the open-weight text-to-music model MiniMax-Music3 on August 13
- Using only lyrics (with section tags) and a structured caption as inputs, it generates a complete song up to 5 minutes long in a single pass, output as 32kHz 16-bit stereo
- Weights, inference code, and three deployment paths — SGLang-Omni, diffusers, and ComfyUI — were released the same day, making it ready for immediate deployment
- 공개일
- 2026년 8월 13일, 오픈웨이트 텍스트투뮤직 모델
- 출력
- 최대 5분 완곡, 32kHz 16비트 스테레오 WAV, 단일 생성
- 핵심 구조
- 8B Global LLM + 0.6B Local LLM(Hybrid-LM) + 2.4B 플로우매칭 + 123M Flow-VAE
- 토크나이저
- 8계층 RVQ, 시맨틱 코드북 16384개 + 어쿠스틱 코드북 7개(각 1024개)
- 실행 경로
- SGLang-Omni(GPU 2장) · diffusers(24GB 이하, 오프로딩 시 8GB) · ComfyUI(FP16/INT8)
- 라이선스
- 상업적 이용 허용, UI에 'MiniMax-Music3' 표시 의무, 연매출 2000만달러 초과 시 서면 승인 필요
- 베이스 모델 불일치
- 모델카드·라이선스는 Qwen3-8B, 미니맥스 리서치 포스트는 Qwen3.5-8B로 표기
A full 5-minute track from a single page of lyrics
Making a song usually means going through composition, arrangement, vocals, and mixing as separate steps. MiniMax-Music3, released by MiniMax on August 13, compresses this process into a single generation pass. Feed it two inputs — lyrics with section tags and a short description — and it outputs a complete track up to five minutes long as a 32kHz, 16-bit stereo WAV file. Weights, inference code, and three deployment paths were all released on launch day, meaning it arrived ready for actual deployment rather than as a research demo.
MiniMax is a Chinese company that trains its own text, image, video, and voice models in-house; its video brand Hailuo and its character chat app Talkie both belong to the same company. On August 3, it open-weighted its video model H3, demonstrating that 15-second video clips could be generated even on a laptop with 8GB of VRAM, and earlier this month it also released the coding agent MiniMax Code 2.0. The music model is the next step in this sequence.
Just two inputs: lyrics and caption
MiniMax-Music3 takes two inputs. The first is lyrics, with each line tagged using section markers such as [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], and [Outro] to specify the song's structure. The second is a structured caption, split into three parts: Global Metadata (genre, mood, etc.), Vocal Details, and Arrangement. A short description alone is enough, since MiniMax's accompanying caption-rewriting agent skill can expand brief sentences into this three-part format.
Two LLMs and flow matching working underneath
The training tokenizer uses 8-layer residual vector quantization (RVQ). The first semantic codebook, with 16,384 entries, captures the song's core structure, while the remaining seven acoustic codebooks, each with 1,024 entries, fill in fine-grained sound detail. During generation, an 8B-scale Global LLM predicts the first codebook frame by frame to establish long-term structure, while a 0.6B Local LLM fills in the remaining codebooks within each frame. MiniMax calls this combination of the two models Hybrid-LM.
The interesting part is the synthesis stage. Rather than decoding the discrete tokens directly, the final hidden states of both LLMs are combined and fed into a 2.4B-scale flow-matching module, which is then reconstructed into latent space by a 123M Flow-VAE inherited from MiniMax Speech. At inference time, the discrete tokenizer decoder itself is never loaded at all. However, documentation differs on the base checkpoint: the model card and license documents state that the Global LLM was initialized from Qwen3-8B, while MiniMax's research blog states Qwen3.5-8B, so the exact version has not yet been confirmed.
How to try it
Three deployment paths are documented. The reference server, SGLang-Omni, requires two GPUs, with one handling Qwen3-based RVQ autoregressive generation and the other handling flow matching and decoding. The diffusers modular pipeline runs under 24GB of VRAM at full precision, dropping to about 22GB with CPU offloading and down to 8GB with leaf-level group offloading. ComfyUI hosts a native Text to Music template built from FP16/INT8 weights repacked by Comfy-Org.
| Deployment path | Requirements | Notes |
|---|---|---|
| SGLang-Omni | 2 GPUs | Reference server, roles split across GPUs |
| diffusers | Under 24GB (8GB with offloading) | Modular pipeline |
| ComfyUI | FP16/INT8 repacked weights | Native template provided |
To actually make a track, first download the model and inference code from Hugging Face or GitHub, then choose one of the three paths above depending on your GPU specs. Then input lyrics with section tags plus the three-part structured caption, and a complete song under five minutes comes out as a WAV file. If you only have a short description, expanding it into the proper format first using the caption-rewriting skill tends to produce more stable results. Use cases where per-track API costs were previously a bottleneck for bulk generation — such as game background music, ad sound branding, background scoring for short-form video, sketch tracks for composers, and mood-based playlist generation — are cited as the model's main areas of application.
What to know before commercial use
The MiniMax-Music3 Community License permits commercial use, but with conditions attached. Products must prominently display the name "MiniMax-Music3" in their UI, and organizations whose combined annual revenue from products using this model exceeds $20 million must obtain separate written approval from MiniMax. Anyone hosting generation services for third parties must also build and maintain their own safeguards against copyright-infringing outputs.
Editor's take
What stands out in this release isn't the track length but how the pipeline was unified into one pass. Most AI composition tools to date have handled melody generation, vocal synthesis, and mixing separately. MiniMax-Music3 generates discrete tokens but then discards them at the decoding stage, feeding the hidden states of both LLMs directly into flow matching instead. That design reflects a judgment that discrete representations are useful for learning structure but become a bottleneck for producing natural-sounding audio. Recalling how the same company's video model H3 unified video, audio, and lip-sync into a single inference pass, MiniMax appears to be applying the same strategy of collapsing a three-stage pipeline into one stage across modalities, now including music.
Hands-on experience with open-weight music generation models has tended to reach the same conclusion. Short loops or background sounds come out convincingly, but full tracks with vocals often show awkward transitions between sections or slurred pronunciation. This is likely why this model adds two layers of control — section tags and the three-part caption. Still, needing a 24GB-class GPU for comfortable operation remains a barrier for individual creators, and it remains to be seen in practice how much speed the group-offloading option that brings requirements down to 8GB actually sacrifices.
For domestic game studios, ad agencies, or short-form content teams, this is worth testing now for demo tracks or background music replacement. However, organizations with annual revenue exceeding $20 million should review the UI display requirement and the separate approval process in advance. The requirement to build one's own copyright-infringement safeguards is also not something to take lightly. For SaaS-style business models that resell generation services to third parties, this clause could translate into real operational costs.
Within a few weeks, the ComfyUI community will likely produce a flood of workflows built around this model. Given the pace at which MiniMax has been open-weighting models across modalities — video, coding, and now music — the next step is likely an integrated creative tool combining all three.
