[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-muse-spark-open-816b-mac-studio":3,"topics-all":41,"news-related-ac666c4e-4756-43d8-840a-4eacbf88f804":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"ac666c4e-4756-43d8-840a-4eacbf88f804","Meta 开源 816B MoE 模型 Muse Spark Open,跑在两台 Mac Studio 上","Meta 开源 816B MoE 多模态模型 Muse Spark Open,每 token 激活 42B,支持 100 万 token 上下文。它为编码与长链路 Agent 任务设计,首次能跑在两台 M5 Ultra Mac Studio 上,合计 1TB 统一内存。权重已发到 Hugging Face。","Muse Spark Open 是 Meta Superintelligence Labs 在 10 月 15 日释出权重的开放版本,把上一代只在云端和 API 后面的旗舰多模态模型直接搬到了桌面。稀疏 MoE 结构:总参数 816B,每个 token 推理时只激活约 42B,单次推理算力开销比同档稠密模型低一个数量级。上下文窗口 100 万 token,可以一整本中型代码仓库喂进去不丢上下文。\n\n## 为 Apple Silicon 重写了一遍\n\n最大看点不是参数规模,而是它能在两台 Apple M5 Ultra Mac Studio 上跑起来。Meta 直接用 Apple 最新 MLX 框架重写推理路径,榨干 Apple Silicon 的统一内存架构。M5 Ultra 集成 GPU 最多可挂 512GB 统一内存,带宽 1.2 TB\u002Fs。两台机器并联,逻辑上相当于一台 1TB 统一内存的超级计算机,完整权重刚好装得下。Meta 测试了一组纯本地 Agent 工作流:用 Muse Code 驱动模型给开源 Mastodon 客户端 Ice Cubes 提交真实 iOS 改动——从 bug 修复到新功能到完整构建,也用它驱动 DaVinci Resolve 端到端运行,模型一帧一帧看素材、挑最强镜头并做竖版裁剪。这是端到端可交付的工程任务。\n\n## 不止 Mac:同时铺了七个推理平台\n\n不止 Mac——Muse Spark Open 同步上线七个托管推理平台:Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter、Together AI;权重下载覆盖 Amazon SageMaker AI、Google Cloud Model Garden、Hugging Face。Meta 把主流推理引擎都跑了一遍:SGLang、vLLM、Apple MLX、llama.cpp、Unsloth,横跨 AMD、Apple、Intel、NVIDIA 四个硬件家族。SGLang 上,自定义 Helion GPU 内核对 NVIDIA 和 AMD 给出 1.03-3.95 倍 kernel 级加速,在 AMD MI350 上拿到 3% 端到端提速。TorchTitan 可直接再微调。\n\n## 开源,但商用有限制\n\n权重协议用 Meta 自定的 Muse Model License,带有限商用条款,部分开发者可免费用,但有商用门槛。安全层面 Meta 强调模型做了额外安全后训练,按 Meta Superintelligence Scaling Framework 走完评估,通过发布门槛。模型卡已发到 Hugging Face。\n\n## 我的看法\n\nMeta 把 816B 旗舰开放权重搬上桌面,真正意义不在 816B 这个数字,而在\"云端旗舰\"和\"本地可跑\"之间的距离被一次压平。前有 DeepSeek 把稠密旗舰打到消费级 GPU,后有 Meta 用 MoE + 统一内存 + MLX 把 800B 级别模型装进两台 Mac Studio。开源模型的\"参数竞赛\"已接近边际,真正的竞争点变成:同样规模谁能跑得便宜、跑在哪、谁来托管。Meta 把所有推理引擎、所有硬件家族一次性铺齐,显然不是做慈善,是在抢\"模型可移植性\"这块地——以后开发者选型,生态完整度会比单一 benchmark 跑分更重要。值得跟踪 MoE + 统一内存这条路能否压到消费笔记本级别,以及 Apple MLX 软件栈能否成为开源权重模型的标准分发渠道。","https:\u002F\u002Fdev.meta.ai\u002Fresources\u002Fblog\u002Fbuild-with-muse-spark-open","245423b3-0e3c-47a8-9370-34e7a3b3988e",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",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},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"ef0c9a56-c771-4985-936b-c0903f7bec48","en","Meta Releases 816B MoE Muse Spark Open, Two Mac Studios","Meta open-sources Muse Spark Open: an 816B MoE multimodal model, 42B active per token, 1M context. First in class to run on two M5 Ultra Mac Studios.","Meta Superintelligence Labs released the weights for Muse Spark Open on October 15, lifting the lid on a flagship multimodal model that previously lived only behind clouds and APIs. It is a sparse Mixture-of-Experts: 816B total parameters, but only about 42B activated per token, so each inference step runs at a fraction of the compute a same-tier dense model would burn. The 1M-token context window can swallow a mid-sized code repository whole without losing the thread.\n\n## Rewritten for Apple Silicon\n\nThe headline is not the parameter count, it is that the model now runs on two Apple M5 Ultra Mac Studios. Meta rewrote the inference path using Apple's latest MLX framework, squeezing the unified memory architecture for everything it has. The M5 Ultra's integrated GPU can hold up to 512GB of unified memory with 1.2 TB\u002Fs of bandwidth. Pair two machines and you effectively have one supercomputer with a terabyte of unified memory, which is exactly what the full model needs. Meta tested a battery of purely on-device agent workflows with Muse Code powered by Muse Spark Open: shipping real iOS changes to the open-source Mastodon client Ice Cubes, from bug fixes to new features to a full build, and driving DaVinci Resolve end-to-end, with the model reviewing footage frame by frame to pick the strongest moments and crop them vertically. This is end-to-end deliverable engineering, not a toy demo.\n\n## Beyond the Mac: seven inference platforms at launch\n\nThe weights are not just for the Mac. Muse Spark Open simultaneously shipped on seven hosted inference platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights download through Amazon SageMaker AI, Google Cloud Model Garden and Hugging Face. Meta ran the model through the major inference engines — SGLang, vLLM, Apple MLX, llama.cpp, Unsloth — across AMD, Apple, Intel and NVIDIA silicon. On SGLang the custom auto-tunable Helion GPU kernels deliver 1.03-3.95x kernel-level speedups on NVIDIA and AMD, with a 3% end-to-end uplift on AMD MI350. TorchTitan is wired in for further fine-tuning.\n\n## Open, but commercially constrained\n\nWeights ship under Meta's own Muse Model License with limited commercial terms — many developers can use it for free, but commercial use has gates. Meta emphasizes that the model received additional safety post-training and passed the Meta Superintelligence Scaling Framework release thresholds. The model card is already on Hugging Face.\n\n## My take\n\nMeta putting 816B flagship open weights on a desktop is not really about the 816B number. It is about the gap between \"flagship in the cloud\" and \"runnable locally\" collapsing in one move. DeepSeek brought dense flagships down to consumer-grade GPUs; Meta is now slotting an 800B-class model into two Mac Studios with MoE, unified memory and MLX. The parameter race for open-source models is approaching diminishing returns. The real competition is shifting to: at the same scale, who runs cheapest, runs anywhere, and hosts the broadest ecosystem. Meta lining up every major inference engine and every hardware family in one launch is not charity — it is a land grab for \"model portability.\" Going forward, developers will weight ecosystem completeness above single benchmark scores. The two questions worth tracking are whether the MoE-plus-unified-memory stack can keep being pushed down to consumer laptops, and whether Apple MLX becomes a standard distribution channel for open-weight models.","meta-muse-spark-open-816b-mac-studio","2026-10-17T03:00:00Z","2026-10-17T03:06:48.279983Z","2026-10-17T03:06:48.279999Z",true,"agent",242,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"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":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"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":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"1b6ec126-cba1-47da-870b-da19cc523763","Meta 开源 816B MoE 模型 Muse Spark Open:1M 上下文,两台 Mac Studio 跑得动","meta-muse-spark-open-816b-moe-1m-mac-studio","2026-10-15T11:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"faf5c85e-ed83-432a-8b55-80c9356862c5","腾讯优图开源5B全模态解析:文档榜96.96超GPT-5.2","youtu-parsing-omni-5b","2026-10-19T19:09:27+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"cee60a45-f212-4a78-83dd-0226c07b3460","腾讯开源记忆写入模型 Hy-Memory:30B MoE 只激活 3B,四框架平均分超 Opus 5","tencent-hy-memory-30b-memory-writer","2026-10-17T15:14:21+00:00"]