要点
What matters
Meta AI Research 于 10 月 15 日发布 Muse Spark Open 权重，称其为原生多模态、816B 总参数/42B 激活参数、1M token 上下文的 agentic 与 coding 模型，并列出多个托管推理伙伴与权重下载渠道。原文描述的 Apple Silicon 桌面方案是两台 M5 Ultra Mac Studio。对希望把创作素材、客服知识或营销流程放在自有环境中的团队，开放权重改变的是部署选项，而不是把基础设施、许可、访问控制、评测和本地运维自动变简单。
Meta AI Research released Muse Spark Open weights on October 15, describing a natively multimodal 816B-total/42B-active model with a 1M-token context for agentic and coding work, plus multiple hosted-inference partners and weight-download channels. The Apple Silicon desktop arrangement described in the post is two M5 Ultra Mac Studios. For teams hoping to keep creative assets, support knowledge or marketing workflows in their own environment, open weights change the deployment options—not make infrastructure, licensing, access control, evaluation or local operations automatically simple.
范围与事实边界
Scope and limits
All benchmark, safety and speed statements are Meta’s reports under its stated methods. The model is released under a Muse Model License with limited commercial terms; the announcement does not establish that every commercial use, hardware stack, hosting partner, country or workflow is permitted, economical or suitable. A 1M-token context is a model capability statement, not a guarantee that a local system can securely supply, retain or reason correctly over a million tokens.
All benchmark, safety and speed statements are Meta’s reports under its stated methods. The model is released under a Muse Model License with limited commercial terms; the announcement does not establish that every commercial use, hardware stack, hosting partner, country or workflow is permitted, economical or suitable. A 1M-token context is a model capability statement, not a guarantee that a local system can securely supply, retain or reason correctly over a million tokens.
编辑工作台
Editorial workbench
在下载前写部署预算：权重与量化版本、显存/内存、吞吐目标、功耗、推理引擎、许可证、素材存放位置、访问身份、日志、漏洞更新和删除路径。用一个脱敏且可回退的小任务与托管方案并行比较，而不是先把完整客户库搬进去。产出是一页自托管准入评审。
Before downloading, write a deployment budget: weights and quantization, VRAM/RAM, throughput target, power, inference engine, licence, asset location, access identity, logs, security updates and deletion path. Compare one de-identified, reversible task in parallel with a hosted route instead of moving a whole customer corpus first. The deliverable is a one-page self-hosting admission review.