Meta AI 研究页集中展示 Muse Spark 1.3、SAM 3 等最新模型成果
AI Research
- 1Muse Spark 系列两个月内连续更新，1.3 版于 2026 年 9 月 2 日发布
- 2Muse Glimmer 开放权重，针对消费级硬件上的常驻本地智能体工作流优化
- 3SAM 3 支持用文本和视觉提示检测、分割并追踪图像或视频中任意物体，同时发布 SAM 3D
- 4安全框架覆盖网络安全、生化安全与失控风险，将为每个前沿模型发布 Preparedness 报告
Meta 用一页研究门户同时推进本地开放权重智能体、视觉基础模型与成体系安全评估，能力说法均为官方口径。
- 影响：Muse Glimmer 开放权重并适配消费级硬件，为本地运行常驻智能体提供了新选项；各模型的安全与评估报告可作为选型尽调的官方材料；页面未给出性能数据和商用价格。
- 参考：可考虑把 Muse Glimmer 列入本地智能体试点候选清单，先在消费级硬件上验证工作流表现；值得把 Muse Spark 系列的安全报告纳入模型采购评估材料。
- 依据：模型名称、发布日期与安全框架要点均取自 Meta 官方研究页，能力描述为官方口径，暂无独立验证。
Muse Your personal AI agent
AI RESEARCH
Our latest AI advancements
Introducing Muse Glimmer
Open weights and optimized for always-on local agent workflows on consumer hardware.
RESOURCES
Muse Image and Muse Video
The first media generation models developed by Meta Superintelligence Labs.
Muse Spark 1.1
Read more about our newest and most capable model and our path to personal superintelligence.
Safety and Preparedness Report
How Muse Spark performs across safety, security, and reliability—and the decisions behind its deployment.
Muse Spark ReportMuse Spark (Contemplating Mode) ReportMuse Spark 1.1 Evaluation Report
WHAT'S NEW
### Introducing Muse Spark 1.3 SEP 2, 2026
### Introducing Muse Code and Muse Spark 1.2 AUG 5, 2026
### Introducing Muse Spark 1.1 JUL 9, 2026
### Introducing Muse Image and Muse Video JUL 7, 2026
PERCEPTION
Segment Anything 3
Our most advanced segmentation model, now with 3D, audio and more capabilities
With SAM 3, you can use text and visual prompts to precisely detect, segment and track any object in an image or video.
SAM 3D
SAM 3D enables precise reconstruction and analysis of 3D people and objects, providing new opportunities for spatial understanding and applications.
SAM Audio
A more conversational Meta Al means you can ask anything, anytime and get helpful answers wherever you are.
CORE LEARNING & REASONING
DINOv3
Our most powerful self-supervised algorithm and vision foundation model
DINOv3 scales self-supervised learning to train a powerful, more versatile model
More from Meta's FAIR Team
Meta Motivo
A first-of-its-kind behavioral foundation model for embodied humanoid virtual agents.
Video Seal
A state-of-the-art, open-source model for video watermarking.
Movie Gen
The most advanced family of media foundation AI models empowering immersive storytelling.
Seamless Communication
AI research models that enable more natural, authentic communication across languages.
AI Chemistry
Building AI systems solving most important chemistry and material problems for Meta and the world.
Show More
SAFETY
Alignment and preparedness
As we chart the path towards superintelligence, AI tools can deeply understand your world and help you get things done faster. With that, reliability, security and user protections are more important than ever. Read more
What is the Advanced AI Scaling Framework?
This framework sets the standards and safeguards we hold our most capable models to — covering how we identify, assess and address the most severe risks before deployment. It's built around comprehensive safeguards, rigorous testing, ongoing research and a level of transparency that goes further than industry norms.
What risks does the framework focus on?
The framework is focused on evaluating the most severe areas of potential risks for advanced models including cybersecurity threats, chemical and biological security and safeguarding against loss of control. As we build more agentic AI experiences, we've expanded assessments to evaluate how AI performs with greater autonomy and whether controls around that behavior work as intended.
Core priorities include: comprehensive safeguards, rigorous testing, ongoing research to identify new risks and mitigation techniques.
How do you test your models for safety?
We stress-test across a wide range of scenarios using automated and human-led evaluations, including adversarial probing. Models are evaluated before and after mitigations are applied to confirm safeguards work in real-world conditions — and only deployed when they meet our framework thresholds.
How has your safety approach evolved?
We've moved beyond rules-based systems to teach models why something is safe — not just what is allowed. By translating guidelines across child safety, response quality and handling different viewpoints into testable principles, models can reason through situations rigid rules might never have anticipated.
What is a Preparedness Report?
A document covering how each frontier model performs across safety, security, accuracy, helpfulness and user protections. We'll publish one for every frontier model — including risk assessments, evaluation results, methodology, known limitations and deployment reasoning.
Try experimental demos
How Meta is applying cutting-edge AI research to real-world interactions.
A state-of-the-art, open-source model for video watermarking
Create video cutouts and effects with a few clicks
Join us in the pursuit of what’s possible with AI
来源：Meta AI：Research Blog · ai.meta.com