Meta Superintelligence Labs has launched Muse Spark 1.1, a multimodal reasoning model built for agentic tasks with substantial improvements in tool use, coding, and computer interaction. The model operates in "Thinking" mode within the Meta AI app and on meta.ai, while developers gain access through the new Meta Model API now in public preview. This release accompanies the debut of Muse Image, advancing Meta's vision of personal superintelligence that can plan, orchestrate, and execute across applications.

Muse Spark 1.1 manages a 1 million token context window, actively compacting history while preserving critical steps for extended workflows. It orchestrates multi-agent systems as both main agent and subagent, delegating parallel execution and escalating when needed. The model demonstrates sophisticated computer-use behavior — writing scripts for automation, clicking for direct interaction, and batching actions — adapting to evolving interfaces across applications like Facebook Marketplace listing creation from smartphone video.

Coding performance shows large gains on real-world tasks involving enterprise codebases. On the Meta Internal Coding Bench, the model significantly improves over its predecessor and competes with leading alternatives. It supports agentic coding setups with planning mode, goal conditioning, and context compaction. Multimodal strengths include visual-to-code generation, ultra-descriptive captioning, and grounded perception-action loops where the model inspects visual and audio inputs while operating computers on users' behalf.

Safety evaluations under the Advanced AI Scaling Framework confirm Muse Spark 1.1 operates within safe margins across chemical, cybersecurity, and loss-of-control risk categories, with strong resistance to jailbreaks and prompt injection. Early partners including Replit, Cline, and Box validate its enterprise readiness. The combination of million-token context, full multimodal support, and OpenAI-compatible API positions it as a complete agentic foundation for production workloads.