Introducing Muse Spark 1.1

Meta Superintelligence Labs pushed out Muse Spark 1.1 today, and the pitch is less about raw chat smarts and more about getting things done. This is a multimodal reasoning model built to act as an agent — planning tasks, delegating to subagents, and operating computers with minimal hand-holding. It slots in right behind this week's Muse Image launch, and Meta is framing both as steps toward what it calls personal superintelligence, its branding for AI that actually executes on your behalf rather than just talking at you.

The headline upgrade is agentic orchestration. Muse Spark 1.1 can run as a main agent that gathers context, builds a plan, and farms out execution to parallel subagents, or it can operate as one of those subagents itself, knowing when to escalate a problem back up the chain. Meta says it zero-shot generalizes to unfamiliar tools, MCP servers, and custom skills, and it actively manages a 1-million-token context window — remembering earlier steps, retrieving old work, and compacting history so it doesn't lose the plot on long jobs. In one demo, the model handled a dinner-party order and adjusted on the fly when new information showed up mid-task, without anyone prompting it to.

Coding is where Meta claims the biggest jump over the original Spark. The model reportedly handles large, messy enterprise codebases better, fixes complex bugs, and manages sizable code migrations. It's also been tuned to work inside popular agentic coding setups — supporting planning mode, subagent delegation, and context compaction — rather than forcing developers into a Meta-only workflow. Internally, Meta says it's now competitive with leading rivals on its own coding benchmark, and researchers are already using it to evaluate other models, which is a neat bit of eating your own cooking.

On the multimodal side, Spark 1.1 leans into tasks where seeing and acting have to happen together — captioning video, generating code from visual references, or in one showcased example, watching a smartphone video of a product and turning it into a finished Facebook Marketplace listing by itself. Meta also says it ran the model through its Advanced AI Scaling Framework and found it within safe margins on chemical, cyber, and loss-of-control risk categories, with better resistance to jailbreaks and prompt injection than before.

The model is live now in Thinking mode in the Meta AI app, and for the first time developers can reach it through a new Meta Model API in public preview. Early partners — Replit, Cline, Box, and the OpenClaw Foundation among them — are already talking it up as a serious agentic coding tool, which suggests Meta wants this positioned less as a novelty demo and more as infrastructure other companies build on top of.