Meta opens Muse Spark 1.1 to developers through new Model API

The multimodal reasoning model is available in Meta AI and meta.ai, with developer access now in public preview.

Meta has opened developer access to Muse Spark 1.1 through the public preview of its Meta Model API.

Meta has opened developer access to Muse Spark 1.1, a new multimodal reasoning model from Meta Superintelligence Labs, through the public preview of its Meta Model API.

The July 9 release gives developers access to the upgraded model for agentic AI, coding, computer-use and multimodal workflows. Muse Spark 1.1 is also available now in "Thinking" mode in the Meta AI app and on meta.ai.

The update follows Meta’s release of Muse Image and places Muse Spark 1.1 in a growing group of AI models built not only to respond to prompts, but to plan work, use tools, operate software and coordinate tasks across longer workflows.

Meta says the new Model API allows teams to start building with Muse Spark 1.1, which it describes as an upgrade over the original Muse Spark in performance, efficiency and tool use.

Built for longer agentic workflows

Meta says Muse Spark 1.1 is designed for agentic tasks that require planning and orchestration across external apps and services.

The model can work with native tools, Model Context Protocol servers and custom skills, according to Meta. It has also been trained to support multi-agent systems, where a main agent gathers context, plans a task and delegates parts of the work to parallel subagents.

That architecture is aimed at reducing the time needed to complete complex projects. Meta says Muse Spark 1.1 can act as the main agent or as a subagent, depending on the workflow.

The model also has a context window of 1 million tokens. Meta says this allows Muse Spark 1.1 to keep track of earlier actions, retrieve information from previous parts of a task and compact context while preserving critical steps for later use.

Those capabilities are likely to be watched by teams building AI assistants for education, workforce training, student support, software development and internal operations, where long-running tasks can involve multiple files, systems and decision points.

Computer use, coding and multimodal work

Meta says Muse Spark 1.1 improves on computer-use workflows that span multiple applications and changing information.

Rather than treating every task as a sequence of manual desktop actions, the model has been trained to decide when scripting is faster and when direct interface use is simpler. Meta says it can also generate batches of actions rather than proceeding one click at a time.

In one example, Meta showed Muse Spark 1.1 organizing a dinner party and updating an order when new context changed the task.

Coding is another focus of the release. Meta says Muse Spark 1.1 performs better than Muse Spark on real-world software tasks involving large codebases, including bug diagnosis, feature implementation and code migration.

The model supports agentic coding setups with planning mode, goal conditioning, subagent delegation and context compaction. In an OpenCode demo, Muse Spark 1.1 built a chat web app, used automated screenshots to identify visible failures, traced the issues to relevant code, implemented fixes and validated the changes.

Meta says its own developers and researchers are using Muse Spark 1.1 for internal coding and model development workflows. The company also says the model improved on its primary internal coding evaluation, Meta Internal Coding Bench, and was competitive with leading alternatives.

Muse Spark 1.1 also handles visual and audio input, according to Meta. The company says the model can support visual-to-code artifact generation, detailed image and video captioning, and workflows where perception and action are combined.

In a Facebook Marketplace example, the model used smartphone video to extract useful product photos, reason about the item and operate a browser to create a listing on a user’s behalf.

Early partners and safety testing

Meta says Muse Spark 1.1 has been tested before deployment under its Advanced AI Scaling Framework, including evaluations for chemical and biological risk, cybersecurity and loss of control.

The company says its evaluations show Muse Spark 1.1 operated within safe margins across those frontier risk categories. Meta also says the model showed resistance to direct jailbreaks, indirect attacks from untrusted data, prompt injection and developer-prompt attacks, while reducing hallucination rates and sycophancy.

Several early partners have commented on the model’s developer and enterprise use cases.

Amjad Masad, CEO of Replit, says: "What’s most impressive about Muse Spark is how much it packs into one model: massive million-token context, full multimodal support (images, video, PDFs), built-in search with citations, strong reasoning, top-tier coding abilities (particularly frontend and design), structured output, and parallel tool calling — all in a clean OpenAI-compatible package. A complete agentic foundation."

Saoud Rizwan, CEO of Cline, links the release to agentic coding workloads: "Meta is clearly building for serious agentic coding – strong tool use at a price point that makes it viable to run real coding workloads at scale. That combination is rare, and it’s exactly why we wanted Cline developers to have access early."

Dave Morin, OpenClaw Foundation, adds: "Muse Spark 1.1 is an awesome model for running agents. Fast, powerful, and fun with OpenClaw.”