Muse Spark 开启

Muse Spark 开启
Muse Spark Open

原始链接: https://research.meta.ai/blog/muse-spark-open

Meta发布了Muse Spark开源，这是一个针对个人超算优化的开放权重AI模型。它原生支持Apple Silicon，通过MLX运行，两台Mac Studio M5 Ultra联合提供1TB统一内存和1.2 TB/s带宽，可完全在桌面端处理大型模型。发布内容包括自定义的Helion GPU内核，集成于SGLang中，可在NVIDIA/AMD显卡上实现高达3.95倍的内核级加速，并在AMD MI350上实现3%的端到端性能提升。安全的后训练确保符合Meta的超智能扩展框架，该模型采用Muse模型许可证，广泛支持非商业用途。现已在Hugging Face上线，用户可使用PyTorch的TorchTitan进行自定义，或原生在Mac上运行。该发布展示了实际的代理工作流案例，包括iOS开发和DaVinci Resolve视频处理，并邀请社区在Hugging Face论坛分享。

黑客新闻新 | 过去 | 评论 | 提问 | 展示 | 招聘 | 提交登录 Muse Spark Open，迄今为止美国最强的开放权重模型 ( meta.ai ) 72 分由 meetpateltech 1 天前 | 隐藏 | 过去 | 收藏 | 1 条评论帮助 dang 1 天前 [–] 评论已移至 https://news.ycombinator.com/item?id=50107379 。回复考虑申请 YC Winter 2027 批次！申请开放至 11 月 2 日。指南 | 常见问题 | 列表 | API | 安全 | 法律 | 申请加入 YC | 联系搜索：

Meta 发布了 Muse Spark Open，一个 8160 亿参数（420 亿激活）的开放权重模型，定位为最强的美国开放权重选项。Hacker News 评论者指出，它在基准测试上落后于 Qwen 3.8 Flash Next 和 GLM 5.3 Flash 等模型，同时在本地运行需要非常昂贵的硬件——大约两台配备约 1TB 统一内存的 Mac Studio M5 Ultra 系统。讨论还将其与 DeepSeek、GLM 和 Meta 的 Muse 产品线进行比较，质疑基准测试的可靠性，并强调许可限制：它并非完全开源，大型提供商必须获得单独的商业协议。有些人认为 Meta 更广泛的 Muse 智能体生态系统是一种消费者 AI 策略，而其他人则质疑其粘性和相关性。总体而言，这次发布被视为美国开放权重 AI 的进步，但并非对中国模型的明显领先者。

原文

Key Benchmarks

For more details about our evaluations, see the evaluation methodology report.

Optimized for Apple Silicon

Leveraging the latest version of MLX, Muse Spark Open was optimized to run on Mac Studio with M5 Ultra, and together they’re the best way to run a powerful model on your own desktop.

Apple’s unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model. And the chip’s 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory.

We tested a range of agentic workflows running entirely on-device, using Muse Code powered by Muse Spark Open, including shipping real iOS changes to Ice Cubes (an open-source Mastodon client) — bug fixes, new features, full build — and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.

Available Now

We’re sharing this work because we want everyone to have access to the best models. The open-weight version of Muse Spark ships with custom, auto-tunable Helion GPU kernels in SGLang, delivering 1.03-3.95× kernel-level speedups on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350.

Safety and Preparedness

Releasing open-weight models is part of how we bring the benefits of personal superintelligence to everyone, and part of how safety work advances across the field.

Open weights also carry their own considerations, and Muse Spark Open has additional safety post-training for that reason. The results demonstrate strong safeguards across risk categories measured.

We evaluated the model in line with our Meta Superintelligence Scaling Framework, covering the most severe risks, and it met the thresholds our framework requires for release. Detailed results are in the Muse Spark Open model card, with a full evaluation report available soon.

License

We’re releasing the weights under the Muse Model License with limited commercial terms, which allow many developers to use the model for free.

Get Started

You can customize Muse Spark Open for your use case by leveraging PyTorch’s TorchTitan to further tune the model. To run it on Apple silicon, use MLX, which pulls the weights directly from Hugging Face.

Muse Spark Open is available now through Hugging Face: https://huggingface.co/collections/meta-models/muse-spark-open

MLX Adapter for running Muse Spark Open on Mac Studio.

Hosted Inference Partners: Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter.

Built or published something with Muse Spark Open? Share with us on the Hugging Face Forums — we’d love to see what you create.