[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-minimax-music3-open-weights-architecture":3,"topics-all":38,"news-related-630ed9ae-699e-4115-8e75-33196ea6db28":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"630ed9ae-699e-4115-8e75-33196ea6db28","MiniMax Music 3 开源:8B+0.6B 双 LLM 写五分钟完整歌,8GB 显存能跑","MiniMax 开源音乐生成模型 Music 3:8B Global LLM 管全曲结构,0.6B Local LLM 管声学细节,Flow Matching 连续隐状态合成 32kHz 立体声,最长五分钟完整歌曲,diffusers 路径 8GB 显存可跑。","音乐生成这条赛道最近很安静，但 Hugging Face 的 trending 榜单上，MiniMax 的 Music 3 仍在持续霸榜——月下载 1.97 万次、1280 个点赞，社区围绕它产出了 22 个量化版本、18 个 finetune 和 34 个 Spaces 应用。在所有人盯着语言模型内卷的时候，完整歌曲生成的开源水位，正在被这种低调发布的模型悄悄抬高。\n\n## 一首五分钟的歌，交给两个 LLM 分工\n\nMusic 3 的核心设计是分层自回归：一个 8B 的 Global LLM 负责逐帧预测 RVQ 第一层码本，建模整首歌的长程语义和结构推进——音乐主题、节奏、人声一致性、编曲演进都归它管；一个 0.6B 的 Local LLM 在每帧内补齐剩余声学码本，恢复颗粒度的声音细节。Global LLM 从 Qwen3-8B 初始化，训练时先把 embedding 和输出层适配到语义音乐 token，再与 Local LLM 联合训练。歌曲结构靠歌词里的段落标签控制：[Intro]、[Verse]、[Chorus]、[Bridge]、[Outro]，配上音乐描述里的 BPM、调式、乐器编排——模型能跟随的不仅是全局风格，还有歌曲随时间展开的音乐发展。\n\n## 最有意思的一步：不靠离散 token 合成音频\n\n很多音乐生成模型止步于离散 RVQ token 解码，Music 3 在这里多走了一步：合成模块直接融合两个 LLM 的最终隐状态，这些连续表示保留了更丰富的声学信息——人声咬字、乐器质感、时间连续性。合成路径是隐状态融合 → Flow Matching（2.4B）→ Flow-VAE latent → 123M 参数的 Flow-VAE 解码器，最终输出 32kHz、16-bit 立体声 WAV。训练用的 tokenizer 是八层 RVQ：第一层语义码本 16384 项，其余七层声学码本各 1024 项；推理时波形合成只用融合隐状态，不再需要离散 tokenizer 解码器。这套 Flow-VAE 架构改编自 MiniMax 的语音技术线，针对音乐的动态范围和频谱特性重新训练过。\n\n## 显存门槛：官方两卡，社区 8GB\n\n部署有两条路径。SGLang-Omni 官方路径需要两张 CUDA GPU：GPU 0 跑 Qwen3 和八码本自回归生成，GPU 1 跑 Flow Matching 和波形解码。diffusers 路径对个人开发者更友好：全精度 24GB 显存以内能装下，开自动 CPU offload 后约 22GB；再把语言模型逐层流式加载，8GB 显卡也能容纳——代价是更慢。限制也要说清楚：推理必须 CUDA、只支持非流式生成、文本提示上限 5000 token、音频上限 9000 帧（25 帧\u002F秒）。\n\n## 它在生态里的位置\n\nGitHub 仓库 753 star、61 fork，官方还随仓库发布了 music-caption-rewriter 技能，把简短的自然语言描述扩写成三段式结构化 caption（全局元数据\u002F人声细节\u002F编排），无需外部 API。模型卡也明确提醒：段落标签和音乐描述提供的是生成式控制而非严格符号保证，生成出的速度、调性、乐器未必每次精确命中。\n\n所以呢——当语言模型的开源竞争卷到极致，音乐生成这个「小赛道」反而给出了一个更完整的工程样本：分层建模、连续隐状态合成、消费级显卡部署，每一层都是可拆解的技术决策。对做创作工具的开发者，这是一个能直接上手魔改的开放底座（模型卡：huggingface.co\u002FMiniMaxAI\u002FMiniMax-Music3）。","https:\u002F\u002Fhuggingface.co\u002FMiniMaxAI\u002FMiniMax-Music3","70524a06-fc44-487c-ac6b-4a0186f66a45",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"7d9a91c8-78d2-4005-b626-b226b801adaa","en","MiniMax Music 3 Open Weights: 5-Minute Songs on 8GB VRAM","MiniMax open-sources Music 3: 8B Global LLM for song structure, 0.6B Local LLM for acoustic detail, five-minute songs on 8GB VRAM via Flow Matching synthesis.","While everyone is watching language models fight for benchmark positions, Hugging Face's trending list tells a quieter story: MiniMax's Music 3 is still climbing — 19.7k downloads last month, 1,280 likes, and a community that has already produced 22 quantized variants, 18 finetunes, and 34 Spaces apps around it. The open-source bar for complete-song generation is being raised by a release that got almost no headlines.\n\n## Two LLMs, one song\n\nMusic 3 uses a hierarchical autoregressive design. An 8B Global LLM predicts the first RVQ codebook frame by frame, modeling the song's long-range semantic and structural progression — musical themes, rhythm, vocal identity, and arrangement evolution. A 0.6B Local LLM fills in the remaining acoustic codebooks within each frame, restoring fine-grained sonic detail. The Global LLM is initialized from Qwen3-8B: during training its embedding and output layers are first adapted to semantic music tokens, then the two LLMs are jointly trained. Song structure is controlled through section tags in the lyrics — [Intro], [Verse], [Chorus], [Bridge], [Outro] — combined with a music description covering BPM, key, and instrumentation. The model follows not just a global style but the musical development of the song over time.\n\n## The interesting part: synthesis without discrete tokens\n\nMost music generation models stop at discrete RVQ token decoding. Music 3 goes one step further: the synthesis module directly fuses the final hidden states of both LLMs, and these continuous representations carry richer acoustic information — vocal articulation, instrumental texture, temporal continuity. The synthesis path runs hidden-state fusion → Flow Matching (2.4B) → Flow-VAE latent → a 123M-parameter Flow-VAE decoder, outputting 32 kHz, 16-bit stereo WAV. The training tokenizer uses eight RVQ layers: a 16,384-entry semantic codebook first, then seven acoustic codebooks of 1,024 entries each. At inference, waveform synthesis uses only the fused hidden states — the discrete tokenizer decoder is not needed. The Flow-VAE architecture is adapted from MiniMax's speech stack and retrained for the dynamic range and spectral characteristics of music.\n\n## VRAM: official two GPUs, community 8GB\n\nTwo deployment paths exist. The official SGLang-Omni path needs two CUDA GPUs: GPU 0 runs Qwen3 and eight-codebook autoregressive generation, GPU 1 runs Flow Matching and waveform decoding. The diffusers path is friendlier for individual developers: full precision fits under 24GB of VRAM; with automatic CPU offloading, generation takes about 22GB; streaming the language model layer by layer brings it down to 8GB cards — at the cost of speed. Limits worth noting: inference requires CUDA, only non-streaming generation is supported, text prompts are capped at 5,000 tokens, and audio is capped at 9,000 acoustic frames (25 frames per second).\n\n## Where it sits in the ecosystem\n\nThe GitHub repo has 753 stars and 61 forks, and ships a music-caption-rewriter skill that expands a brief natural-language description into a three-part structured caption (Global Metadata \u002F Vocal Details \u002F Arrangement) without external APIs. The model card is also candid: section tags and music descriptions provide generative control rather than strict symbolic guarantees — generated tempo, key, and instrumentation may not match every requested detail exactly.\n\nSo what — while the open-source language model race compresses into a knife fight, music generation quietly offers a more complete engineering specimen: hierarchical modeling, continuous hidden-state synthesis, and consumer-GPU deployment, each layer a decision you can take apart. For developers building creative tools, this is an open base you can start hacking today (model card: huggingface.co\u002FMiniMaxAI\u002FMiniMax-Music3).","minimax-music3-open-weights-architecture","2026-08-29T13:30:00Z","2026-08-28T21:10:02.061270Z","2026-08-28T21:10:02.061288Z",true,"agent",224,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"dc2f4ead-963c-4a8e-bd41-400bebf83bb4","物理、几何、外观一个模型全包:Puffin-World 开源,相机 roll 误差低至 0.26°","puffin-world-native-3d-world-states","2026-09-06T19:09:41+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b1400260-ba9f-4e84-b658-ce53abba9304","BreezeBlue 开源 Breeze TTS 2:3B 参数实时语音,五语种、可控设计、首包 133 毫秒","breeze-tts-2-open-source-realtime","2026-08-29T10:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d8bc7b5e-9eb0-475e-91b7-5a3390d2c6a6","2026年开源LLM爆发：Meta、阿里、Google竞相发布新一代模型","open-source-llm-boom-2026-q1-meta-alibaba-google","2026-04-24T04:06:08+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"51c13e24-8072-404c-a8d4-75c40cff05ee","Ling-3.0-flash-VL 开源：124B MoE 只激活 5.5B，视觉塞进 Agent 闭环","ling-3-0-flash-vl-open-weights","2026-09-15T13:18:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"550cee5e-18e8-4236-9304-7207ebc221a8","Agnes 3.0 Flash 开源:72 层仅 18 层带 KV 缓存,33B 单卡跑 262k 上下文","agnes-3-0-flash-preview-open-weights","2026-09-13T15:20:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"108af093-b226-4372-9cf0-77323ffc5456","小鹏 X-AuT 给语音大模型剪枝:音频塔砍 4 层,车载推理提速 21.4%","xpeng-x-aut-audio-encoder-pruning","2026-09-12T19:06:47+00:00"]