Meta unveiled the latest version of its multimodal Spark AI model on Thursday, Spark 1.1, positioning it as a major step forward for agentic AI—systems that don’t just answer questions but take actions on a user’s behalf. The company claims the model delivers big gains in computer use, coding, and multimodal understanding, marking its push to re-enter the AI race after a year of reorganization and lagging behind rivals like OpenAI and Anthropic. The model was released by Meta’s Superintelligence Labs, led by former Scale AI CEO Alexandr Wang.

Spark 1.1 is built around a multiagent system: a main agent creates a plan and delegates tasks to parallel subagents, which use tools and escalate back when needed. Meta says the model actively manages a 1-million-token context window by remembering actions, retrieving earlier work, and compacting context to handle more complex tasks without hitting limits. On the coding front, Spark 1.1 promises improvements in diagnosing bugs, adding features to enterprise systems, and performing large code migrations. It can also take control of a user’s computer for tasks, maintaining context across long sessions and navigating unfamiliar interfaces with minimal input.

The model’s multimodal abilities let it see and hear to complete tasks—ingesting imagery and audio to generate ultra-descriptive captions, then interpreting on-screen content to take action. Meta also launched a public preview of its Meta Model API, making Spark 1.1 available to developers, and it’s now live in “Thinking” mode within the Meta AI app and on meta.ai. The company touts extensive safety evaluations under its Advanced AI Scaling Framework, claiming strong resistance to jailbreaks and prompt injection.

Source: CNET