开发者实测本地音乐生成模型MiniMax-Music3

MiniMax-Music3这款本地音乐生成模型发布了，我尝试体验了一下（截至2026年8月，测试环境为RTX 5090）。

提示词越简短，输出的音乐音符往往越少。官方演示的提示词基本上每首要用到600～800个词。我本来觉得这根本不是人力能写出来的量……结果发现官方GitHub里附带了叫music-caption-rewriter的功能。大概就是让我们用LLM把短提示词展开成要求的格式吧。

我让Claude Code读取了这个功能，让它帮我写提示词。我觉得词数少生成的简约节拍完全没问题。但对于喜欢旋律扎实的日本流行乐或摇滚的人来说，体验就不好说了……

它达不到Suno那种只输入几个词，比如“情绪流行乐”，就能生成出一首听得过去的作品的水平。也很难说它更像Suno的哪个模型（硬要说的话是v5.5），整体反而更接近ACE-Step 1.5 XL。

它不太适合像Suno那样硬把完全不相关的风格融合到一起。日语人声的表现相当扎实。视频生成模型H有时候会把汉字读成中文发音，但目前来看Music3的发音错误很少，混有日语和英语的歌词也能分清楚演唱。

但这款本地模型没有Suno具备的翻唱、延长这类功能。截至2026年8月，它只能从文本和歌词一次性生成音乐。

生成时间方面，模型加载完成后，时长基本和歌曲实际时长差不多（60秒的歌大约需要50秒，2分半的带人声歌曲大约需要3分钟）。即使使用相同随机种子，只要改变生成时长，就会输出完全不同的歌曲。

附件是使用相同歌词、相同随机种子的对比试听：A版是直接把给Suno写的提示词原样输入的版本，B版是展开成官方格式的版本。我还额外加入了用和A相同的提示词生成的Suno v4.5版（C）和v5.5版（D）供参考。但要说明的是，虽然Suno版（C）（D）用了和A相同的提示词，但每次点生成都会产出不同的歌曲（应该是随机种子变了），所以并不是公平的测试环境。

――A. C. D. 和输入给Suno的提示词相同（150词）――
Vocals: Japanese female idol style, straight-tone (flat) delivery with minimal vibrato, bright head-voice. Bright, modern J-Pop / Idol Pop fused with retro City-Pop funk-disco sparkle, 115 BPM, E♭ major, Punchy 80s kick & snare with gated verb, groovy slap bass, clean wah-guitar chops, FM pad chords, brassy synth stabs, sparkly saw arps, hand-claps, glitter FX, short risers, talkbox solo in bridge, Structure: Intro → Verse 1 8b → Pre-Chorus 4b → Big Chorus 8b with crowd chants "Hey!" "Woo!" → Verse 2 → Pre-Chorus → Chorus → Funky instrumental break → Half-step key change → Double Chorus → A-capella tag "la-la, la-la, la“ fade-out, Vocals: cheerful Japanese female idol trio, bright head-voice, tight harmony stacks, playful call-and-response, whispered ad-libs, Lyrics theme: neon summer nights, first crush, chasing dreams, carefree optimism, Mood: upbeat, nostalgic yet fresh, danceable, kawaii sparkle, Production: crisp side-chained mix, shiny reverb, wide stereo, master at −8 LUFS

――B. Music3官方格式（用music-caption-rewriter展开，678词）――
Global Metadata Basic Attributes: bpm is 115. key is Eb, and scale is major. J-Pop / Idol Pop / Retro City-Pop Funk-Disco. Global Emotional Progression: The track opens with a glittering, anticipatory sparkle - neon signs flickering on at dusk - and settles into verses that skip along with carefree, first-crush optimism. Each pre-chorus gathers itself over four bars of rising claps and a short riser, and every chorus bursts wide open: big, bright, danceable joy with the crowd shouting along. A funky instrumental break turns the energy playful and strutting, a talkbox trading licks with the band; then a half-step key change lifts the double final chorus one notch higher into pure celebration, before the song dissolves into a tender a-capella tag that fades like the end of a perfect summer night. Application Scenarios & Imagery: Neon summer nights in the city, an idol trio on an outdoor festival stage, first crushes and chased dreams, convertible rides past glowing arcades. Penlights, hand-claps from the crowd, glitter falling under stage lights - nostalgic yet fresh, kawaii sparkle everywhere. Sonics & Production Profile: Crisp, modern side-chained dance mix with a wide stereo image and shiny reverb. Punchy 80s-style kick and a big gated-reverb snare drive the retro backbone; the master is loud and polished with a bright, sparkling top end. Vocals sit up front, doubled and clean; claps and chants spread wide; every transition marked with short risers and glitter FX.

Vocal Details Vocal Gender & Timbre: A cheerful Japanese female idol trio. Bright head-voice timbres, cute yet precise, blending into tight three-part stacks - youthful, polished and charming. Vocal Style: Straight-tone (flat) delivery with minimal vibrato and crisp Japanese diction. Verses are playful and rhythmic; pre-choruses tighten and rise; choruses are sung full-out by all three in bright unison, every syllable landing on the beat. Harmony/Backing Vocals: Tight harmony stacks bloom in every chorus with a third above on the hook words. Playful call-and-response between the lead and the other two members runs through the verses; crowd-style gang chants answer the big choruses; whispered ad-libs flicker between lines. The song closes with an a-capella group tag repeating a soft la-la phrase, fading out. Vocal FX: Clean pitch-perfect idol sheen, light doubling, shiny plate reverb in the choruses, drier and closer in the verses, and a talkbox voice appearing only in the instrumental break as a solo instrument.

Arrangement Primary: Brassy synth stabs punch the chord accents and answer the vocal lines - the instrumental signature - joined by sparkly saw-wave arpeggios that shimmer across the stereo field in every chorus like glitter. Secondary: A groovy slap bass locks with the kick in funk-disco syncopation; clean wah-guitar chops ride tight sixteenths; warm FM pad chords hold the harmony; hand-claps stack the backbeat; a talkbox lead takes the solo in the break. Groove & Foundation Progression: A punchy 80s-flavored disco-funk beat: tight kick four-to-the-floor with syncopated pushes, a big gated snare on the two and four, crisp hats and layered hand-claps. The intro sparkles in with arps and a riser over the groove. Verse one rides eight bars of slap bass, wah chops and pads with the drums lighter; the four-bar pre-chorus stacks rising claps and a short riser; the chorus hits eight bars at full width with brass stabs, saw arps and crowd chants. Verse two and the second pre-chorus follow the same shape with small additions, into the second full chorus. Then the funky instrumental break: slap bass and wah guitar strut while the talkbox solos over the band. A quick riser lifts a half-step key change into the double final chorus - two passes, everything at maximum plus extra claps and arps an octave up. The band cuts, and the trio”s a-capella la-la tag repeats softly, fading to silence. Textures & Spatial FX: Glitter FX and short risers into every section, gated-reverb tails on the snare, wide stereo arps panning like stage lights, a sparkle-dust transition into the key change, and a gentle studio air under the a-capella fade-out.

这是指定风格生成纯器乐节拍的作品。808贝斯陷阱、Afrobeats生成得都很像模像样，所以它对近期流行的曲风表现力更强？G-Funk它好像完全没理解😂 我个人觉得哪怕只能无限生成这类节拍就已经很开心了。

MiniMax-Music3好像不认识艺术家名、制作人名这类信息。

Text Encoder有完整版（bf16）和轻量版（pruned int8），热身完成后的生成速度差距大约是20%（轻量版更快。60秒的歌曲完整版50秒/轻量版40秒）。

官方仓库和演示页面在这里。