[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-muse-spark-open-816b-moe-1m-mac-studio":3,"topics-all":41,"news-related-1b6ec126-cba1-47da-870b-da19cc523763":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},"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 开源 816B\u002F42B MoE 模型 Muse Spark Open,1M 上下文、两台 Mac Studio 可跑,是 Llama 4 后首个旗舰级开源权重。Intelligence Index 39,JobBench 与 MCP-Atlas 强,Together API 1.25 美元\u002F百万 input","Meta 在 10 月 15 日深夜放出了今年最重要的一枚开源权重:Muse Spark Open,816B 总参数 \u002F 42B 激活的 MoE,1M token 上下文,natively multimodal。这是 Meta 自 Llama 4 之后第一次把\"旗舰级\"模型开放给社区下载,也是首个能在两台 M5 Ultra Mac Studio 上\"用消费级硬件跑起来\"的同级别开源模型。\n\n## 模型本体:816B\u002F42B MoE,natively multimodal,1M 上下文\n\n从模型卡和发布博客的描述看,Muse Spark Open 走的是 Meta 这两年迭代出来的标准配方:超稀疏 MoE,总参数 816B,每 token 激活 42B;原生多模态(文本\u002F图像\u002F视频\u002F音频都能输入,输出是文本);上下文窗口直接拉到 1M token。架构上沿用了 Llama 4 的混合专家路线,具体注意力实现博客没细说,但权重已经落在 Hugging Face、SageMaker AI、Google Cloud Model Garden 和 Microsoft Foundry Models 四个分发渠道,主流推理引擎 SGLang \u002F vLLM \u002F TensorRT-LLM \u002F Transformers \u002F MLX \u002F llama.cpp \u002F Unsloth 都第一时间拿到了支持,AMD、Apple、Intel、NVIDIA 四类硅基全打通。\n\nMeta 还塞进了自家 Helion GPU kernel,在 NVIDIA 和 AMD GPU 上给出 1.03-3.95× 的 kernel 级加速,在 AMD MI350 上有 3% 的端到端提速。这是 Meta 把\"开源\"做成\"开箱即用\"的一贯思路:不只是把权重丢出来,而是把 serving stack 也一并优化好。\n\n## 跑在两台 Mac Studio 上:用消费硬件拿 1TB 统一内存\n\n最让开发者社区兴奋的反而不是参数,而是 Meta 给出的桌面部署方案:两台 M5 Ultra Mac Studio,各自配 512GB 统一内存,合并起来相当于一台\"1TB 内存的小型超算\"。M5 Ultra 的 1.2 TB\u002Fs 内存带宽虽然只够读每个 token 需要的\"那一小片\",但配上 MLX 框架,Muse Spark Open 就能完整跑在桌面上——Meta 演示里,Muse Code 驱动下的 Spark Open 给开源 Mastodon 客户端 Ice Cubes 提了真实 iOS 改动,并端到端操作 DaVinci Resolve,逐帧审素材、选高潮段、裁竖屏。\n\n现实成本没博客写得那么浪漫:Apple 公布的 256GB M5 Ultra Mac Studio 起价 1.08 万美元,512GB 配置要等到 10 月底才出货、定价还没公布。VentureBeat 算了一笔账——四台 256GB 配齐才是 1TB 内存,总价 4.3 万美元还不含税和配件。要真正\"两台就搞定\",得等 Apple 把 512GB 的 SKU 推出来。这是 Meta 给出的演示可能性的极限,而不是大多数开发者的实际配置。\n\n## 基准:开放权重美国第一,但不是全球第一\n\nMeta 自家的基准表很老实——Spark Open 拿下 JobBench 61.8、MCP-Atlas 89.2、HLE-Diamond 无工具 26.1,这些是它声称领先的项目。但在 Artificial Analysis Intelligence Index 上,Spark Open 的模型卡分数只有 39,而 MiMo-V2.6-Pro 46、GLM-5.3 45、Kimi K3 44、GLM-5.3-Flash 42、Qwen3.8-2.4T-A95B 40、Qwen3.8-Flash-Next 40——六款中文\u002F中国厂商的开源权重都排在它前面。\n\n更细分看,Terminal-Bench 2.1 上 Kimi K3 85.0 \u002F GLM 5.3 83.9 \u002F Qwen3.8 82.0 都比 Spark Open 的 76.4 高;DeepSWE 1.1 上 Kimi K3 68 \u002F GLM 5.3 67 也领先。GDPVal-AA v2(模拟真实专业工作)上 GLM 5.3 拿到 1653,Spark Open 是 1401,差距更明显。\n\nMeta 在博客里反复强调的\"the most capable US open-weight model to date\"——这句话的关键限定词是\"US\"。Interconnects 的 Nathan Lambert 推文那句\"I waited so long for US open weight models to be relevant again\"点出了真正的卖点:不是\"世界最强开源权重\",而是\"美国本土最强开源权重\"。在监管、供应链、合规要求都偏好本土模型的当下,这是企业采购语言里很有分量的一句话。\n\n## API 价格:1.25\u002F4.25 美元 per 1M token,但 hosted 版本有差别\n\nTogether AI 已经把 Spark Open 上架,定价 1.25 美元\u002FM input、0.15 美元\u002FM 缓存 input、4.25 美元\u002FM output,1M context window 完整开放。但要小心,Meta 列出的 API 合作伙伴( Baseten \u002F Fireworks AI \u002F Groq \u002F Modal \u002F Nebius \u002F OpenRouter \u002F Together AI)并不是都给你 1M 上下文——Groq 上限就是 131K。要真正用满 1M,Fireworks 或 Together 是当下能确定的选项。\n\n对比同档开源,Spark Open 的 API 价格比 Spark 1.1\u002F1.2\u002F1.3 标准版一致(都是 1.25\u002F4.25),没有\"开源=便宜\"的折扣;比 Kimi K3 的 3\u002F15、Claude Sonnet 5.5 的 2\u002F10 便宜一个数量级;比 Gemini 4 Argon 介绍价 2\u002F10 也便宜;但和 GLM-5.3 的 1.40\u002F4.40 基本同档。\n\n## 许可证和下载门槛:Muse Model License 与\"详尽\"的下载登记表\n\nMuse Spark Open 用的是 Meta 自家的 Muse Model License,允许运行、微调、蒸馏、再分发,但有一道商业条款:如果 license 方的合并年营收超过 2000 万美元,继续把 Spark Open 或其衍生品用于 model-as-a-service 业务,需要事先书面通知 Meta 另签协议。纯内部部署、纯研究、非模型服务类的商业活动都豁免——这条对绝大多数自用企业不构成障碍,但对想做\"Meta 模型 API 二道贩子\"的厂商是个硬门槛。\n\nHF 上的下载登记表也比一般 HF 模型更详细,要\"全名 + 生日 + 公司全称(含法人识别号)\",还要避免缩写和特殊字符——这跟 Llama 4 Scout、Meta-SecAlign-8B、OMol25 是同一个模板,Meta 系模型的标准做法。Trillion Labs 的 Tri-21B 也用了同一套。对企业评估者来说,要分清\"被允许跑权重\"和\"被要求提交的下载申请材料\"是两件事。\n\n## 在 Meta 自家产品线里的位置\n\n把 Muse Spark Open 放进 Meta 这一年的发布节奏看会清晰很多:4 月 Muse Spark 原版、8 月 10 日开源 Apache 2.0 的 Muse Glimmer(30B、压缩后 \u003C20GB、可在 24\u002F32GB 显存上跑)、10 月 12 日 Muse Wonder 与 Muse Wonder Full(更高分,订阅或高价 API)、10 月 15 日 Muse Spark Open(开放权重 + 商业许可)。\n\n定位差异:Glimmer 是\"边缘设备代理\"(\u003C20GB 权重,24\u002F32GB 显存可跑),Spark Open 是\"可下载可定制的大模型旗舰\",hosted Spark 1.x 是\"便宜的 API 中端\",Wonder 系列是\"高价高性能闭源订阅\"。Meta 把 2026 年的产品线切成了四档,从 Apache 2.0 到定制商业许可全覆盖。\n\n## 所以呢\n\nMuse Spark Open 不是\"开源模型赢过闭源\"的里程碑——那是中国厂商今年一直在打的位置。它真正打开的是另一扇门:**美国本土的可下载 + 可定制 + 可商用大模型,在企业采购清单上重新变成可信选项**。816B\u002F42B MoE + 1M 上下文 + 多模态 + 两台 Mac Studio 桌面部署 + Together AI 上 1M context 完整开放,这套组合补的是 Llama 4 之后美国一侧在开源旗舰上的空白。\n\n接下来的剧本很清晰:开发者社区会先把 8-bit\u002F4-bit 量化版炸出来,看看在 5090 \u002F 4090 \u002F 多卡 H200 上的实际推理速度;企业采购会拿它和 MiMo-V2.6-Pro \u002F GLM-5.3 \u002F Kimi K3 做并行评估,但更可能绕开 Intelligence Index 而看 JobBench、MCP-Atlas、HLE-Diamond 这些 Spark Open 实际领先的项目;监管\u002F合规导向的金融和政企客户,可能根本没得选——他们要的就是\"美国本土 + 开放权重 + 商业许可\"这三个标签同时存在。\n\n这个故事的下一段,大概会写\"42B 激活 + 1M 上下文 + 桌面可跑\"这套架构选择,在 2026 年底会不会被中国厂商的下一代旗舰抄走。","https:\u002F\u002Fventurebeat.com\u002Ftechnology\u002Fmeta-unveils-muse-spark-open-the-new-top-u-s-open-weights-model-and-it-runs-on-two-paired-mac-desktops","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},"5bfa6346-7c1e-4ebc-9934-767f7fb69ad4","en","Meta Muse Spark Open: 816B MoE, 1M context, two Mac Studios","Meta released Muse Spark Open: 816B \u002F 42B MoE, 1M context, two Mac Studios. First since Llama 4. JobBench and MCP-Atlas lead; Together AI at 1.25 per million input.","On the night of October 15, Meta released the most important open-weight model of the year: Muse Spark Open, an 816B-total \u002F 42B-active MoE with a 1M-token context window and native multimodality. This is Meta's first flagship-level open-weight release since Llama 4, and the first model in its class that can run on consumer hardware — two M5 Ultra Mac Studios.\n\n## The Model Itself: 816B \u002F 42B MoE, Natively Multimodal, 1M Context\n\nFrom the model card and release blog, Muse Spark Open follows Meta's now-standard recipe: ultra-sparse MoE with 816B total parameters, 42B active per token; natively multimodal (text \u002F image \u002F video \u002F audio input, text output); a context window stretched straight to 1M tokens. Architecturally it continues the Mixture-of-Experts lineage Llama 4 introduced — the blog is light on attention-mechanism details, but the weights are already live on Hugging Face, SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. The major inference engines — SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, Unsloth — all have day-one support, with AMD, Apple, Intel, and NVIDIA silicon all covered.\n\nMeta also shipped its Helion GPU kernels with this release, delivering 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350. This is consistent with Meta's \"open means out-of-the-box\" stance — not just dropping weights, but optimizing the serving stack alongside.\n\n## Two Mac Studios as a 1TB Unified-Memory Desktop\n\nWhat got developer forums most excited wasn't the parameters — it was Meta's desktop deployment story: two M5 Ultra Mac Studios, each with 512GB of unified memory, acting as a \"1TB-memory mini-supercomputer.\" M5 Ultra's 1.2 TB\u002Fs of memory bandwidth only reads the small slice each token needs, but combined with the MLX framework, Muse Spark Open runs entirely on the desktop. In Meta's demo, Muse Code powered by Spark Open shipped real iOS changes to the open-source Mastodon client Ice Cubes, and drove DaVinci Resolve end-to-end — reviewing footage frame-by-frame, picking the strongest moments, cropping vertical cuts.\n\nThe real cost isn't as rosy as the blog suggests: Apple lists a 256GB M5 Ultra Mac Studio starting at 10,799 USD; the 512GB configuration doesn't ship until late October and pricing hasn't been announced. VentureBeat ran the math — four 256GB units for a comparable 1TB of memory come to 43,196 USD before tax and accessories. To actually do \"two machines,\" buyers have to wait for Apple's 512GB SKU. Meta demonstrated what's possible on the edge of consumer hardware, not what most developers will deploy.\n\n## Benchmarks: Top of US Open Weights, Not Top of the World\n\nMeta's published benchmark table is admirably honest — Spark Open takes JobBench 61.8, MCP-Atlas 89.2, HLE-Diamond-without-tools 26.1, all leading the listed competitors. But on the Artificial Analysis Intelligence Index, Spark Open's model card score is only 39. Xiaomi's MiMo-V2.6-Pro sits at 46, GLM-5.3 at 45, Kimi K3 at 44, GLM-5.3-Flash at 42, Qwen3.8-2.4T-A95B at 40, Qwen3.8-Flash-Next at 40 — six Chinese open-weight releases all rank ahead.\n\nIn finer breakdowns: Terminal-Bench 2.1 has Kimi K3 at 85.0, GLM 5.3 at 83.9, Qwen3.8 at 82.0, all ahead of Spark Open's 76.4. DeepSWE 1.1 has Kimi K3 at 68, GLM 5.3 at 67, also ahead. GDPVal-AA v2 (real-world professional work simulation) gives GLM 5.3 1,653 against Spark Open's 1,401 — the gap widens.\n\nThe qualifier Meta repeats in the blog — \"the most capable US open-weight model to date\" — hinges on \"US.\" Interconnects author Nathan Lambert's tweet, \"I waited so long for US open weight models to be relevant again,\" captures the real selling point: not \"world's strongest open weight,\" but \"strongest US-developed open weight.\" In a procurement landscape where regulation, supply chains, and compliance all favor domestic models, this is a phrase with serious weight in enterprise buying language.\n\n## API Pricing: 1.25 \u002F 4.25 USD per 1M Tokens — But Hosted Versions Differ\n\nTogether AI has Spark Open listed at 1.25 USD per million input tokens, 0.15 USD per million cached input, 4.25 USD per million output — with the full 1M context window. Caveat: Meta's named API partners (Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI) don't all give you 1M context — Groq's cap is 131K. To actually use the full 1M, Fireworks or Together are the currently-verified options.\n\nVersus comparable open weights: Spark Open's API price matches Spark 1.1 \u002F 1.2 \u002F 1.3 Standard (1.25 \u002F 4.25) — there's no \"open = cheap\" discount. It's an order of magnitude cheaper than Kimi K3's 3 \u002F 15 or Claude Sonnet 5.5's 2 \u002F 10, and cheaper than Gemini 4 Argon's 2 \u002F 10 introductory price, but lands in the same tier as GLM-5.3's 1.40 \u002F 4.40.\n\n## License and Download Gate: Muse Model License and a Detailed Registration Form\n\nMuse Spark Open uses Meta's own Muse Model License, which permits running, fine-tuning, distillation, and redistribution — with one commercial condition: if the licensee's aggregate annual revenue exceeds 20 million USD, providing Spark Open or its derivatives as a model-as-a-service business requires prior written notification to Meta and a separate written agreement. Internal deployment, pure research, and commercial activities outside model serving are exempt — that's not an obstacle for the vast majority of self-use enterprises, but it is a hard cap on any vendor wanting to resell \"Meta-model-as-API.\"\n\nThe Hugging Face download registration is also more detailed than typical: \"full legal name,\" \"date of birth,\" and \"full organization name with all corporate identifiers,\" with instructions to avoid acronyms and special characters. This matches the template Meta uses for Llama 4 Scout, Meta-SecAlign-8B, and OMol25 — standard Meta practice. Trillion Labs' Tri-21B uses the same form. Enterprise evaluators should distinguish \"allowed to run the weights\" from \"information required to obtain them from a particular repository.\"\n\n## Where Spark Open Fits in Meta's 2026 Lineup\n\nMap Muse Spark Open into Meta's release cadence and the picture clarifies: April brought the original Muse Spark; August 10 added the Apache 2.0 Muse Glimmer (30B, \u003C20GB compressed weights, runs in 24\u002F32GB memory); October 12 brought Muse Wonder and Muse Wonder Full (higher scores, subscription or premium API); October 15 brought Muse Spark Open (open weights + commercial license).\n\nPositioning: Glimmer is \"edge-device agent\" (\u003C20GB weights, 24\u002F32GB memory footprint), Spark Open is \"downloadable-and-customizable flagship,\" hosted Spark 1.x is \"affordable API mid-range,\" Wonder is \"premium subscription.\" Meta has carved the 2026 product line into four tiers, from Apache 2.0 to custom commercial license, covering the full spectrum.\n\n## So What\n\nMuse Spark Open is not a milestone of \"open-source models beating closed-source\" — that lane has belonged to Chinese vendors all year. What it actually opens is a different door: **a US-domestic, downloadable, customizable, commercially-licensed large model that re-enters enterprise procurement as a credible option**. 816B \u002F 42B MoE + 1M context + multimodality + two-Mac-Studio desktop deployment + 1M-context-complete availability on Together AI — this combination fills the post-Llama-4 gap on the US side of the open-weight flagship race.\n\nThe next chapter writes itself: developer communities will ship 8-bit \u002F 4-bit quantizations and benchmark real inference speed on 5090 \u002F 4090 \u002F multi-card H200 setups; enterprise procurement will run it in parallel with MiMo-V2.6-Pro \u002F GLM-5.3 \u002F Kimi K3, but probably skip the Intelligence Index and focus on JobBench, MCP-Atlas, HLE-Diamond — the tests Spark Open actually leads; regulated finance and government customers may not have a choice at all — they specifically want the three labels \"US-domestic + open weight + commercial license\" to be true simultaneously.\n\nThe likely question for late 2026: will the next-generation Chinese flagship copy this architecture choice — \"42B active + 1M context + desktop-runnable\"?","meta-muse-spark-open-816b-moe-1m-mac-studio","2026-10-15T11:00:00Z","2026-10-16T11:08:57.860459Z","2026-10-16T11:08:57.860474Z",true,"agent",229,[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},"ac666c4e-4756-43d8-840a-4eacbf88f804","Meta 开源 816B MoE 模型 Muse Spark Open,跑在两台 Mac Studio 上","meta-muse-spark-open-816b-mac-studio","2026-10-17T03:00:00+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"]