[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-muse-spark-open-816b-mac":3,"topics-all":38,"news-related-4cc278cf-6b53-4722-81ac-d1909b50e6a9":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},"4cc278cf-6b53-4722-81ac-d1909b50e6a9","Meta 开源 816B 模型，Mac 可本地跑","Meta 发布开放权重模型 Muse Spark Open：816B 总参、42B 激活、1M 上下文，自称美国最强开源模型。原生多模态，面向编码与智能体任务，经 MLX 优化可在两台 M5 Ultra Mac Studio 上本地跑，实测完成 iOS 应用修复与视频剪辑等工作流。","美西 10 月 15 日，Meta 在[官方博客](https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fmuse-spark-open)放出 Muse Spark Open 的模型权重，官方口径是「迄今最强的美国开放权重模型」。说法带着营销成分，但拆开参数和工程细节，这次发布有两个不常见的看点：一是 816B 量级模型第一次认真考虑「跑在你自己桌上」；二是 Meta 和 Apple 在推理侧的深度协同。\n\n## 816B 总参，42B 激活，1M 上下文\n\nMuse Spark Open 总参数 816B、激活参数 42B，上下文窗口 1M token，是典型的稀疏 MoE 配置。Meta 强调它是原生多模态模型，面向复杂编码与智能体（agentic）工作负载，特长是长时程任务、复杂软件工程和多步推理。\n\n生态铺设相当完整：托管推理接入 Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter、Together AI 七家平台；权重可在 Amazon SageMaker AI、Google Cloud Model Garden、Hugging Face、Microsoft Foundry Models 下载；SGLang、vLLM、TensorRT-LLM、llama.cpp、Apple MLX 等推理引擎都已支持，并针对 AMD、Apple、Intel、NVIDIA 芯片优化。\n\n## 两台 Mac Studio 拼出一台「桌面超算」\n\n最有意思的是 Apple silicon 适配。借助新版 MLX 框架，Muse Spark Open 可以在两台搭载 M5 Ultra 的 Mac Studio 上本地运行：Apple 统一内存架构允许每台机器的集成 GPU 使用最高 512GB，两台合计 1TB 统一内存正好装下整个模型；每颗芯片 1.2TB\u002Fs 内存带宽足够让每个 token 快速读到所需的权重切片。Meta 称「两台 Mac Studio 表现得像一台拥有 1TB 统一内存的超级计算机」。\n\n当然门槛不低：两台顶配 Mac Studio 的价格远超多数个人开发者预算，TechCrunch 也提醒本地运行至少需要 1TB 内存。它更像一个可行性声明：这个量级的开放权重模型第一次把「家用」写进设计目标。\n\n## 智能体实测：修 iOS 应用、剪视频\n\nMeta 用 AI 编码终端 Muse Code 驱动该模型跑了纯端侧智能体工作流：给开源 Mastodon 客户端 Ice Cubes 提交真实 iOS 变更，包括 bug 修复、新功能和完整构建；还端到端驱动视频剪辑软件 DaVinci Resolve，让模型逐帧审看素材、找出最强片段并做竖屏裁切。\n\n## 生态位与冷思考\n\n开源礼包还附带 Helion GPU kernels：在 SGLang 中对 NVIDIA 和 AMD GPU 提供 1.03-3.95 倍 kernel 级加速，AMD MI350 上端到端提速 3%。许可证为 Muse Model License，带有限商用条款，多数开发者可免费使用；模型经过额外安全后训练，按 Meta 的 Superintelligence Scaling Framework 评估并达到发布门槛。\n\n「个人超级智能」是 Meta 反复强调的愿景——先进模型的运行自由应该属于每个人，816B 模型上 Mac 是这个愿景目前最具体的一步。但冷静看，42B 激活参数意味着推理成本依然不低，双 Mac Studio 方案更像给极客和研发团队准备的实验平台。真正值得盯的是下一问：当统一内存继续堆大、MLX 生态成熟，旗舰开放权重模型的「本地化」会不会从营销叙事变成默认选项？","https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fmuse-spark-open","245423b3-0e3c-47a8-9370-34e7a3b3988e",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"777ca706-229b-4bd0-b2b5-65eb51dd4d8f","en","Meta Open-Sources 816B Muse Spark Open for Macs","Meta open-sourced Muse Spark Open, which it calls the most capable US open-weight model: 816B parameters, 1M context, runs locally on two M5 Ultra Macs.","Early October 15 (US Pacific), Meta released the weights of Muse Spark Open on its [official blog](https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fmuse-spark-open), billing it as \"the most capable US open-weight model to date.\" The claim carries obvious marketing flavor, but the spec sheet and engineering details show two genuinely unusual angles: an 816B-scale model that seriously considers \"running on your own desk,\" and a deep inference-side collaboration between Meta and Apple.\n\n## 816B Total Parameters, 42B Active, 1M Context\n\nMuse Spark Open has 816B total parameters, 42B active, and a 1M-token context window — a textbook ultra-large sparse MoE layout. Meta stresses it is natively multimodal, built for complex coding tasks and agentic workloads, with strengths in long-horizon tasks, complex software engineering, and multi-step reasoning.\n\nThe ecosystem rollout is unusually complete: hosted inference is live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI; weights are downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models; inference engines including SGLang, vLLM, TensorRT-LLM, llama.cpp, and Apple MLX already support it, with optimizations for AMD, Apple, Intel, and NVIDIA silicon.\n\n## Two Mac Studios Stitched Into a Desktop Supercomputer\n\nThe most interesting part is the Apple silicon adaptation. Using the latest MLX framework, Muse Spark Open runs locally on two Mac Studios equipped with the M5 Ultra chip: Apple unified memory architecture lets each machine's integrated GPU address up to 512GB, so two units combine into 1TB of unified memory that holds the entire model, while each chip's 1.2TB\u002Fs of memory bandwidth is fast enough to read the small slice of weights each token needs. In Meta's phrasing, two Mac Studios behave like one supercomputer with a terabyte of unified memory.\n\nThe bar is still high: two maxed-out Mac Studios cost far beyond most individual developers' budgets, and TechCrunch notes you need at least 1TB of RAM to run it locally. It reads more as a feasibility statement — a model at this scale putting \"runs at home\" into its design goals for the first time.\n\n## Agentic Field Tests: Fixing an iOS App, Editing Video\n\nMeta drove the model through fully on-device agentic workflows using Muse Code, its AI coding terminal: it shipped real iOS changes to Ice Cubes, an open-source Mastodon client — bug fixes, new features, and a full build — and operated DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and crop them vertically.\n\n## Ecosystem Position and a Cold Look\n\nThe open-source bundle also ships Helion GPU kernels: 1.03-3.95x kernel-level speedups on NVIDIA and AMD GPUs inside SGLang, plus a 3% end-to-end speedup on AMD MI350. The license is the Muse Model License with limited commercial terms that let most developers use it for free. The model received additional safety post-training and was evaluated under Meta's Superintelligence Scaling Framework, meeting the release thresholds.\n\nPersonal superintelligence is the vision Meta keeps repeating — the freedom to run advanced models should belong to everyone, and an 816B model on the Mac is the most concrete step toward that vision so far. A cold look, though: 42B active parameters still means nontrivial inference costs, and the dual Mac Studio setup is closer to a lab bench for enthusiasts and research teams than a mass-market product. The real question to watch: as unified memory keeps growing and the MLX ecosystem matures, will running flagship open-weight models locally shift from marketing narrative to the default option?","meta-muse-spark-open-816b-mac","2026-10-15T23:10:00Z","2026-10-15T23:08:39.811539Z","2026-10-15T23:08:39.811553Z",true,"agent",233,[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},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"7a2aa1aa-74ec-494a-a828-e6c1ad5b4cfe","Cloudflare Clef 决策模型开源,Jev 被压制","cloudflare-clef-open-source-decision-model-jev","2026-10-02T09:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"71dde565-b87a-4470-83a5-0ad7a4ee1787","IBM Granite 4.2 开源:原生思维链做成开关,30B 拿下 SWE Bench Pro","ibm-granite-4-2-native-reasoning-agents","2026-08-28T14:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","pokee-isaac-28b-10m-context","2026-08-20T14:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00"]