Meta 发布开源权重模型 Muse Spark Open,816B 参数、1M 上下文

来源:PyTorch:Blog(RSS) 2026年10月15日 01:55 AIHOT 评分:83 精选
摘要:Meta 发布开源权重模型 Muse Spark Open,已在 Hugging Face 以 Muse Model License 提供。模型总参数 816B、激活 42B,上下文窗口 1M token,面向编码与智能体工作流,原生多模态(图像、音频、文档),并相比 8 月发布的 Muse Spark 增加了安全后训练。
推荐理由:Meta 开放了 Muse Spark 的权重,并给出 Helion 内核与 TorchTitan 微调支持,读者可据此评估其部署与定制路径。
# Meta 发布开源权重模型 Muse Spark Open,816B 参数、1M 上下文 > 原标题:Muse Spark Is Now an Open-Weights Model - 来源:PyTorch:Blog - 发布时间:2026-10-15T17:55:52.000Z - AIHOT:https://aihot.news/items/s80tw5g2v22j1se4bmqejyaq4 - 原文:https://pytorch.org/blog/muse-spark-open/ ## 摘要 Meta 发布开源权重模型 Muse Spark Open,已在 Hugging Face 以 Muse Model License 提供。模型总参数 816B、激活 42B,上下文窗口 1M token,面向编码与智能体工作流,原生多模态(图像、音频、文档),并相比 8 月发布的 Muse Spark 增加了安全后训练。 ## 推荐理由 Meta 开放了 Muse Spark 的权重,并给出 Helion 内核与 TorchTitan 微调支持,读者可据此评估其部署与定制路径。 ## 正文 · 中文译文 今天,Meta 发布了 [Muse Spark Open](https://research.meta.ai/blog/muse-spark-open) 的权重,这是迄今为止能力最强的美国开放权重模型。它今天已在 Hugging Face 上以 Muse Model License 提供。Muse Spark Open 在 SGLang 和 vLLM 推理引擎上进行了优化运行,并针对 NVIDIA、AMD 和 Intel 进行了优化。 基于我们 [8 月发布的](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) Muse Spark,这个开放权重版本增加了额外的安全后训练,使其在没有系统护栏的情况下作为独立模型使用时更能抵御滥用。其结果是该模型成为安全方面的顶级开源选择。Muse Spark Open 是一个 816B 总参数/42B 激活参数的模型,上下文窗口为 1M tokens。Muse Spark Open 专为编码和智能体工作而构建,原生多模态,并与 Muse Spark 1.3 共享相同的基础模型。 Meta 热爱开源,作为 PyTorch 的创造者,我们致力于回馈社区,以确保整个生态系统完全支持 Muse Spark Open。首先,我们集成了 [Helion](https://pytorch.org/blog/helion/),这是我们激动人心的全新 PyTorch 原生高级 DSL,用于高性能且可移植的 ML 内核,以提供高性能运算。此外,[TorchTitan](https://github.com/pytorch/torchtitan) 已经开箱即用地支持 Muse Spark,用于公开的监督微调和 RL,我们很期待看到 PyTorch 社区能把它推进到多远! ## Muse Spark Open Muse Spark Open 是一个针对编码优化的开放权重模型。它专门针对智能体编码工作流进行训练,提供更高的首次尝试准确率、可靠的工具调用,以及对开发者实际使用的代码仓库的深度代码库理解。你可以在我们之前的模型发布博客中阅读更多内容 [此处](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2)。 ### 增强的编码能力 擅长代码生成、调试和代码库理解。基于真实开发者工作流训练——多文件重构、调试会话和长时间运行任务——因此模型能跟上你实际的工作方式。1M 上下文窗口意味着它可以处理你的整个项目,包括复杂依赖、遗留代码和数千个文件,而无需将任务拆分成多个部分 ### 多模态理解 Muse Spark Open 具备音频和音视频感知能力——以及图像和文档理解。作为一个推理模型,它原生且可靠地处理图像、音频和文本输入。 ### 智能体 Muse Spark Open 从一开始就作为智能体模型构建。它规划步骤、行动、观察,并在长时间跨度内自我纠正——而不仅仅是单次工具调用。它经过训练可以浏览网页(包括多模态浏览),并使用智能体框架来:为可分解的工作生成并协调子智能体;以及利用统一于桌面、移动端和浏览器的计算机使用能力。  它基于长时间跨度、多站点任务进行训练,并生成单独的评论,让模型在行动前说明意图——因此你始终可以知道它在做什么以及为什么。 ## 为什么使用 Helion 内核? Helion 是一种 PyTorch 原生的内核编写语言,专为跨异构硬件平台编写高性能内核而设计。在开源生态系统的早期采用之后,Helion 于 2026 年 4 月成为 [PyTorch Foundation](https://pytorch.org/blog/pytorch-foundation-welcomes-helion-as-a-foundation-hosted-project-to-standardize-open-portable-and-accessible-ai-kernel-authoring/) 的一部分。Muse Spark Open 随附了 RoPE、RMSNorm、量化等操作的 Helion 实现,在 NVIDIA GB300 上,相比用硬件特定语言(例如 NVIDIA CUDA 和 CuTe DSL)编写的自定义内核,实现了 1.03 – 3.95 倍的加速。这些相同的内核可直接移植到 AMD,在 AMD Mi350 上取得了类似的性能提升,端到端吞吐量整体提升了 3%。 此外,我们还在 SGLang 中为 Muse Spark Open 推出了 Helion megakernels。这些内核完全用 Helion 编写,将多个操作融合到单个持久内核中,以改善低延迟推理,并且同样可以针对不同的形状和硬件进行自动调优。 用户可进一步针对更广泛的硬件平台和工作负载对 Helion 内核进行自动调优——包括不同的加速器架构(NVIDIA、AMD、Google TPU 等)、量化格式、张量形状和并行策略——而无需为每种配置维护单独的手工调优实现。这些能力共同使实现可移植性能变得更加容易,同时降低了开发和维护专用内核的成本。 ## 使用 TorchTitan 进行后训练 TorchTitan 是一个 PyTorch 原生平台,专为生成式 AI 模型的快速实验和大规模训练而设计。从今天起,TorchTitan 新增了 Meta 的 Muse Spark Open 的模型定义,因此开发者可以根据自己的需求对其进行调优。 TorchTitan 开箱即用地原生支持多轮 SFT 和 LoRA,以及强化学习能力(TitanRL),允许用户对 Muse Spark Open 进行微调。 ## 开始使用 我们分享这项工作,是因为我们希望每个人都能使用最好的 AI 模型。Muse Spark Open 随附了 SGLang 中自定义、可自动调优的 Helion GPU 内核,在 NVIDIA 和 AMD GPU 上实现了 1.03–3.95 倍的内核级加速,在 AMD MI350 上实现了 3% 的端到端加速。你还可以利用 PyTorch 的 [TorchTitan](https://github.com/pytorch/torchtitan/) 进一步调优模型,以满足你的用例需求。 立即开始构建:[https://huggingface.co/meta-models/Muse-Spark-open-816B-A42B](https://huggingface.co/meta-models/Muse-Spark-open-816B-A42B) _Muse Spark 是我们模型系列中的全尺寸模型。我们之前发布了 Muse Glimmer,一个小型稠密模型。今天就开始使用它,借助_ [_ExecuTorch_](https://pytorch.org/blog/fast-ondevice-agentic-ai-with-executorch/)。 文章 [Muse Spark Is Now an Open-Weights Model](https://pytorch.org/blog/muse-spark-open/) 首次发表于 [PyTorch](https://pytorch.org/)。 ## 正文 · 原文 Today, Meta is releasing the weights for [Muse Spark Open](https://research.meta.ai/blog/muse-spark-open), the most capable US open-weights model to date. It’s available today on Hugging Face under the Muse Model License. Muse Spark Open runs optimized on the SGLang and vLLM inference engines, with optimization for NVIDIA, AMD, and Intel. Building on our [August release](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) of Muse Spark, this open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails. The result is a model that is a top open source choice for safety. Muse Spark Open is an 816B total/42B active parameter model with a context window of 1M tokens. Built for coding and agentic work, Muse Spark Open is natively multimodal and shares the same base model as Muse Spark 1.3. Meta loves open source, and as the creators of PyTorch, we’re committed to giving back to ensure the entire ecosystem fully supports Muse Spark Open. First, we’ve integrated [Helion](https://pytorch.org/blog/helion/), our exciting new PyTorch-native high-level DSL for performant and portable ML kernels, to ship high-performance operations. Additionally, [TorchTitan](https://github.com/pytorch/torchtitan) already supports Muse Spark out of the box for supervised fine-tuning and RL in the open, and we’re excited to see how far the PyTorch community can take it! ## Muse Spark Open Muse Spark Open is a coding-optimized open-weight model. Trained specifically for agentic coding workflows, it delivers higher first-attempt accuracy, reliable tool calling, and deep codebase understanding across the repos developers actually work in. You can read more in our previous model release blog [here](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2). ### Enhanced Coding Capabilities Strong at code generation, debugging, and codebase understanding. Trained on real developer workflows — multi-file refactors, debugging sessions, and long-running tasks — so the model keeps up with how you actually work. 1M context window means it handles your entire project including complex dependencies, legacy code, and thousands of files without needing to break your task into pieces ### Multimodal Understanding Muse Spark Open has audio and audio-video perception — alongside image and document understanding. As a reasoning model, it processes image, audio, and text inputs — natively and reliably. ### Agentic Muse Spark Open is built from the ground up as an agentic model. It maps out steps, acts, observes, and self-corrects over long horizons — not just single-shot tool calls. It’s trained to browse the web (including multimodal browsing), and use an agentic harness to: spawn and coordinate subagents for decomposable work; and leverage computer use capabilities unified across desktop, mobile, and browser.  It’s trained on long-horizon, multi-site tasks, and generates separate commentary that lets the model state intent before acting—so you can always know what it’s doing and why. ## Why Helion kernels? Helion is a PyTorch native kernel authoring language designed for writing high-performance kernels across heterogeneous hardware platforms. Following early adoption from the open-source ecosystem, Helion became a part of the [PyTorch Foundation](https://pytorch.org/blog/pytorch-foundation-welcomes-helion-as-a-foundation-hosted-project-to-standardize-open-portable-and-accessible-ai-kernel-authoring/) in April 2026. Muse Spark Open ships with Helion implementations of operations such as RoPE, RMSNorm, quantization, etc., delivering 1.03 – 3.95x speedups compared to custom kernels written in hardware specific languages (e.g. NVIDIA CUDA and CuTe DSL) on NVIDIA GB300. These same kernels were directly portable for AMD, scoring similar performance improvements on AMD Mi350 with an overall 3% improvement in end-to-end throughput. In addition, we also launched Helion megakernels for Muse Spark Open in SGLang. Written entirely in Helion, these kernels fuse multiple operations into a single persistent kernel to improve low latency inference, and can likewise be autotuned for different shapes and hardware. Helion kernels can be further autotuned by users for broader hardware platforms and workloads – including different accelerator architectures (NVIDIA, AMD, Google TPU, etc.), quantization formats, tensor shapes, and parallelism strategies – without maintaining separate hand-tuned implementations for every configuration. Together, these capabilities make it easier to achieve portable performance while reducing the cost of developing and maintaining specialized kernels. ## Post-training  with TorchTitan TorchTitan is a PyTorch native platform designed for rapid experimentation and large-scale training of generative AI models. Starting today, TorchTitan adds a model definition for Meta’s Muse Spark Open , so developers can tune it to their needs.  TorchTitan features native support for multi-turn SFT and LoRA out-of-the-box, as well as reinforcement learning capabilities (TitanRL) that allow users to fine-tune Muse Spark Open. ## Get Started We’re sharing this work because we want everyone to have access to the best AI models. Muse Spark Open 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. You can also customize it for your use case by leveraging PyTorch’s [TorchTitan](https://github.com/pytorch/torchtitan/) to further tune the model. Start Building Today: [https://huggingface.co/meta-models/Muse-Spark-open-816B-A42B](https://huggingface.co/meta-models/Muse-Spark-open-816B-A42B) _Muse Spark is the full size model from our model family. We previously released Muse Glimmer, a small dense model. Start using it today, with_ [_ExecuTorch_](https://pytorch.org/blog/fast-ondevice-agentic-ai-with-executorch/). The post [Muse Spark Is Now an Open-Weights Model](https://pytorch.org/blog/muse-spark-open/) appeared first on [PyTorch](https://pytorch.org/).