Meta releases Muse Spark 1.1 and opens the Meta Model API

Background

Muse Spark 1.1 is an upgrade to the Muse Spark model. Meta described it as a multimodal reasoning model built for agentic tasks, meaning work in which the model takes actions with tools and carries a task through several steps, rather than only answering a single question.^[1] Multimodal means the model handles more than one kind of input, such as text and images.

What happened

On 9 July 2026 Meta released Muse Spark 1.1.^[1] Meta said it improved tool and computer use, coding and multimodal understanding compared with Muse Spark.^[1] Computer use means the model operates a computer's interface, such as clicking and typing, to complete a task.

Key details

Context. Muse Spark 1.1 has a 1-million-token context window. Meta said the window is managed through memory, retrieval and compaction, which means the model keeps the most useful information and summarises or drops the rest as a task grows.^[1] Retrieval here means pulling relevant earlier material back in when it is needed.

Multi-agent work. The model can act as a main agent that delegates work to parallel subagents. It also works zero-shot with new tools, MCP servers and custom skills, which means it can use a tool it has not been specifically trained on, without extra examples. MCP (Model Context Protocol) is a standard way to connect models to outside tools and data.^[1]

Developer access. Developers gained access through the Meta Model API, which was offered in public preview. Early partners included Replit, Cline and Box.^[1] A public preview lets developers try a product before it is generally available.

Consumer use. Muse Spark 1.1 powers Thinking mode in the Meta AI app and on meta.ai.^[1]

Impact

Muse Spark 1.1 extended the Muse series toward tasks that run over many steps, with tools and other agents. Opening the Meta Model API gave outside developers a route to build with the model. Its use in Thinking mode linked the upgrade to Meta's consumer assistant as well.^[1]