Meta on Thursday announced the next version of its multimodal Spark AI model, Muse Spark 1.1, which it calls a significant improvement from its predecessor. The model is specifically designed for agent tasks and promises significant benefits in areas such as computing, coding and multimodal cognition.
Spark 1.1 follows the introduction of the Muse Photoits new tool for generating images, which already has some Instagram users opting out. (You can read how to do that here.)
With the likes of OpenAI again Anthropic Making headlines with new and advanced AI models, Meta’s announcement is a sign that the company is jumping back into the AI race after more than a year behind schedule and extensive restructuring. The model was released by what Meta calls Superintelligence Labs, an outfit led by former Scale AI CEO Alexandr Wang.
Meta also launched a public preview of its Meta Model API, where the new Spark 1.1 will be accessible to developers. The model is now available in “Thinking” mode within the Meta AI app and on meta.ai.
Improved agent skills
Meta says Spark 1.1 can complete complex projects much faster than its predecessor. It is trained in “programming multiagent systems to optimize end-to-end latency,” according to the company’s blog. Agent AI refers to models that can take actual actions for you, rather than providing responses to notifications.
This multiagent system consists of a main agent and subagents. The main agent will execute the program and pass its execution to all the same subagents. Those subagents will stick to their assigned tasks, use available tools and escalate back to the main agent when needed.
Spark 1.1 also manages its 1 million token total window by remembering actions, retrieving information from previous work and combining context in a way that saves important steps needed later, according to the blog. That means you should be able to ask it to do complex tasks without running into the limits of how much the model can handle at once.
Computing and coding
A new model may also be able to control your computer with some jobs, if that’s something you’re willing to try.
Meta says its new AI model has boosted its computing performance across multiple applications. It can maintain context over long periods of time and can navigate in unfamiliar environments with little interaction from the user. It can understand when certain tasks are happening and when to navigate around the computer manually.
On the code front, Spark 1.1 promises significant improvements, especially for tasks involving complex data. It can diagnose and fix bugs, add new features to enterprise-grade systems and perform large code migrations.
It is truly multimodal
The Spark 1.1’s multimodal capabilities allow it to see and hear to complete tasks. This is unlike the existing AI models we’ve seen today, such as the Gemini Live camera experience. It can import images and audio, and provide “more descriptive” captions from sources. Combining this with his computer skills, he can interpret what is on the screen and take action to perform tasks.
Safety
Meta proposes its comprehensive model safety assessment, following the Advanced AI Measurement Framework. The model is said to show strong resistance to jailbreaks, quick injection and other common attacks.
