Muse spark 1.1 arrives as Meta’s new multimodal reasoning model for agents

Meta Superintelligence Labs has introduced Muse Spark 1.1, an update to its multimodal reasoning model built for agentic tasks, with improvements in tool use, computer operation, coding, and image and video understanding. The release comes the same week as Muse Image, another model from the company.

The system can actively manage a context window of one million tokens: it remembers past actions, retrieves information from earlier steps, and condenses previous work while keeping what matters for later tasks. According to Meta, this lets it act as a main agent, gathering context and delegating to subagents, or as a subagent that knows when to escalate a task back.

In benchmarks published by the company, Muse Spark 1.1 scores 88.1 points on MCP Atlas, a large-scale tool-use evaluation, compared to 82.2 for both its predecessor and Opus 4.8, and 75.3 for GPT-5.5. On JobBench, focused on professional tasks, it reaches 54.7 points, ahead of Opus 4.8 (48.4) and GPT-5.5 (38.3). In coding it shows a marked improvement: on Vibe Code Bench v1.1 it jumps from 19.7 to 72.2 points.

In computer-use tasks, the model decides when to automate actions through scripts and when to interact directly with the interface, adapting to information that changes mid-task, such as in an order-management example where it adjusts a message without user intervention.

Meta says it evaluated the model under its internal advanced AI scaling framework, covering chemical, biological, cybersecurity and loss-of-control risks, and states it stays within safe margins, with greater resistance to manipulation attempts and a lower hallucination rate.

The model is available in "Thinking" mode in the Meta AI app and on meta.ai, and for the first time developers can access it through a public preview of the new Meta Model API.

Key points

• Meta introduces Muse Spark 1.1, an update to its multimodal reasoning model for agents
• Improves tool use, computer operation, coding, and multimodal understanding
• Manages a one-million-token context window with memory of past actions
• Can act as a main agent or delegate tasks to subagents
• Outperforms Opus 4.8 and GPT-5.5 on tool-use benchmarks (88.1 on MCP Atlas, 54.7 on JobBench)
• Jumps from 19.7 to 72.2 points on Vibe Code Bench v1.1 in coding
• Evaluated under Meta's safety framework; stays within safe margins
• Available in the Meta AI app, on meta.ai, and through a public preview of the Meta Model API