Meta发布Muse Spark Open多模态MoE模型及权重
🤖 AI Agents Weekly: Muse Spark Open, Gemini Agents Beat Minecraft, OpenCode 2.0, Nemotron 3.5 Super VL, OpenRouter Armor, Claude OpenTelemetry, and More
Meta最新旗舰级开源多模态MoE模型，原生适配智能体场景且支持端侧部署，对Agent开发者极具参考价值，建议重点关注其性能对比与本地运行方案。
In today’s issue:
本期内容：
- Meta releases Muse Spark Open
- Gemini Argon agents beat Minecraft
- OpenCode 2.0 ships Code Mode
- NVIDIA launches Nemotron 3.5 Super VL
- OpenRouter launches Armor for tool calls
- Claude Platform adds OpenTelemetry traces
- DeepSeek ships V4.1-Flash-1017
- Fish Audio opens Drama 3 weights
- Forgetting helps autoresearch agents scale
- Anthropic replaces git with Scion
- Meta 发布 Muse Spark Open
- Gemini Argon 智能体在《我的世界》中胜出
- OpenCode 2.0 上线 Code Mode
- NVIDIA 推出 Nemotron 3.5 Super VL
- OpenRouter 为工具调用推出 Armor
- Claude Platform 增加 OpenTelemetry 追踪
- DeepSeek 发布 V4.1-Flash-1017
- Fish Audio 开放 Drama 3 权重
- 遗忘机制助力自主研究智能体扩展
- Anthropic 用 Scion 替换 git
And all the top AI dev news, papers, and tools.
以及所有顶级 AI 开发新闻、论文和工具。
Top Stories
头条新闻
Muse Spark Open
Meta released the weights for Muse Spark Open, which it calls the most capable US open-weight model to date. It is a natively multimodal Mixture-of-Experts model with 816B total and 42B active parameters and a 1M-token context, built for coding and long-horizon agentic work.
Meta 发布了 Muse Spark Open 的权重，称其为迄今为止最强大的美国开源权重模型。它是一个原生多模态混合专家（MoE）模型，总参数量 816B，激活参数量 42B，上下文窗口为 1M token，专为编码和长周期智能体工作设计。
- Agentic results: The model card reports 89.2 on MCP-Atlas, 61.8 on JobBench, 80.4 on OSWorld Verified, 72.8 on Toolathlon Verified, and 76.4% on Terminal-Bench 2.1 with the Terminus 2 agent. It scores 39 on the Artificial Analysis Intelligence Index, behind GLM 5.3, Kimi K3, and Qwen3.8-2.4T in Meta’s comparison table.
- Runs on a desktop: Meta says it is the first model in its class to run on two Apple M5 Ultra Mac Studios with MLX, and tested it on-device by shipping iOS changes to the Ice Cubes Mastodon client with Muse Code.
- Availability: Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry, with hosted APIs from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. It runs on SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, and Unsloth.
- License: The weights ship under the Muse Model License, whose limited commercial terms let many developers use the model for free.
- Muse Wonder: Earlier in the week, Meta introduced Muse Wonder and Muse Wonder Full, its most capable models for personal and professional agents, with Active Memory and Active Focus. They are available through the Meta Model API and Muse Code.
- 智能体表现：模型卡片显示，在 MCP-Atlas 上得分为 89.2，JobBench 为 61.8，OSWorld Verified 为 80.4，Toolathlon Verified 为 72.8，在使用 Terminus 2 智能体的 Terminal-Bench 2.1 上得分为 76.4%。在 Artificial Analysis Intelligence Index 上得分为 39，在 Meta 的比较表中落后于 GLM 5.3、Kimi K3 和 Qwen3.8-2.4T。
- 桌面端运行：Meta 表示它是同类中首个能在两台配备 MLX 的 Apple M5 Ultra Mac Studio 上运行的模型，并通过使用 Muse Code 向 Ice Cubes Mastodon 客户端推送 iOS 更改，在设备上进行了测试。
- 可用性：权重已上架 Hugging Face、Amazon SageMaker AI、Google Cloud Model Garden 和 Microsoft Foundry，并由 Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter 和 Together AI 提供托管 API。它支持 SGLang、vLLM、TensorRT-LLM、Transformers、llama.cpp 和 Unsloth。
- 许可证：权重采用 Muse Model License 发布，其有限的商业条款允许许多开发者免费使用该模型。
- Muse Wonder：本周早些时候，Meta 推出了 Muse Wonder 和 Muse Wonder Full，这是它们面向个人和专业智能体的最强大模型，具备主动记忆（Active Memory）和主动聚焦（Active Focus）功能。它们可通过 Meta Model API 和 Muse Code 获取。
Blog | Weights | Muse Wonder
博客 | 权重 | Muse Wonder
Read more
阅读全文
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力