Meta's Latest AI Model Puts Agentic Power Front and Center

Meta has rolled out the next generation of its Spark AI family, unveiling Muse Spark 1.1 on Thursday. The company is positioning this multimodal model as a major step forward, especially for agentic tasks - meaning it can actually do things for you rather than just answer questions. The release comes as Meta scrambles to regain momentum in the AI race, after months of playing catch-up to rivals like OpenAI and Anthropic.

The model is built to handle complex, multi-step projects with speed. Meta says Spark 1.1 uses a main agent that breaks down tasks and hands them off to specialized subagents, which can work in parallel, use tools, and escalate back to the lead agent when needed. It also actively manages its 1-million-token memory window by compressing and recalling past actions, so users can tackle longer, more intricate jobs without hitting capacity limits. Beyond planning, the model can take over your computer - navigating unfamiliar interfaces, automating repetitive steps, and keeping context across long sessions. On the coding side, it's designed to diagnose bugs, add features in enterprise systems, and handle large code migrations.

Spark 1.1 is truly multimodal: it can see and hear, ingesting images and audio to generate ultra-detailed captions, and then act on what it interprets on screen. Meta is also opening up a public preview of its Meta Model API, making Spark 1.1 accessible to developers, and it's already live in "Thinking" mode inside the Meta AI app and on meta.ai. The company touts strong safety evaluations under its Advanced AI Scaling Framework, claiming the model resists jailbreaks, prompt injection, and other common attacks.

Source: CNET