Meta describes Muse Spark 1.1 as a multimodal reasoning model designed for agentic applications that require planning, coordination, and interaction with external tools and services. According to the company, the model delivers significant gains over the original Muse Spark in coding, computer use, multimodal understanding, and overall performance efficiency.
A central focus of the update is the model’s ability to coordinate complex, multi-step tasks. Meta said Muse Spark 1.1 can act as a primary agent that gathers context, develops execution plans, and delegates work across multiple subagents running in parallel. It can also function as a subagent, following assigned tasks, using available tools, and determining when additional coordination is needed.
The company said the model supports a context window of one million tokens, allowing it to retain information across extended workflows, retrieve earlier details when needed, and condense context while preserving information required for future tasks.
Meta also highlighted improvements in computer-use capabilities, positioning Muse Spark 1.1 for workflows that span multiple applications and changing information. Rather than relying exclusively on direct interaction with software interfaces, the model determines when scripting or automation is more efficient than manually navigating applications. Meta said the system was trained to generate scripts, interact with interfaces directly when appropriate, and execute batches of actions during longer workflows.
Coding was another major area of focus. Meta said Muse Spark 1.1 performs better on real-world software engineering tasks involving large codebases, including debugging, feature development, and large-scale code migrations. The company added that the model supports common agentic coding workflows such as planning, subagent delegation, context management, and goal conditioning.
Within Meta, developers and researchers are already using Muse Spark 1.1 for software development and model evaluation tasks, according to the company. Meta said the model showed substantial improvement over its predecessor on the company’s internal coding benchmark while remaining competitive with leading alternatives.
The update also expands the model’s multimodal capabilities. Meta said Muse Spark 1.1 can combine visual understanding, audio processing, reasoning, and tool use to complete longer workflows. The company highlighted examples including generating code from visual inputs, producing detailed image and video descriptions, and creating Facebook Marketplace listings by extracting images from smartphone video while operating a web browser on a user’s behalf.
Alongside the product announcement, Meta said it completed safety evaluations under its Advanced AI Scaling Framework before deployment. The company reported that Muse Spark 1.1 remained within its deployment thresholds across evaluated frontier risk categories, including chemical and biological risks, cybersecurity, and loss of control. Meta also said the model demonstrated resistance to jailbreak attempts, prompt injection attacks, and other adversarial techniques while reducing hallucinations and sycophantic behavior.
The launch also broadens developer access to the Muse family. While the original Muse Spark model was previously limited to select partners through a private API preview, Meta is now introducing a public preview of the Meta Model API. According to Meta, some partners already have access, while additional developers can join a waitlist as access expands over time. For now, API availability is limited to Meta’s own platform rather than third-party model marketplaces.
In an interview discussing the release, Meta Superintelligence Labs Chief AI Officer Alexandr Wang described Muse Spark 1.1 as the company’s strongest model yet for agentic workloads and coding. He said the infrastructure supporting the API is operated directly by Meta and characterized pricing as competitive with similar offerings. New API accounts will receive $20 in free credits, after which the company will charge $1.25 per million input tokens and $4.25 per million output tokens.
Wang also said Meta intentionally emphasized coding performance because it directly strengthens AI agents capable of completing broader autonomous tasks. He added that the company trained Muse Spark 1.1 to work with widely used agentic development frameworks to encourage developer adoption.
Although Meta has shifted its recent AI strategy toward proprietary models, Wang said the company remains committed to open source and confirmed that an open-source variant of Muse Spark is in development, though he did not provide a release timeline.
Meta said additional AI models are already in training, signaling that Muse Spark 1.1 is part of a broader roadmap as the company continues expanding its portfolio of agentic AI systems.
This analysis is based on reporting from Meta.
Image courtesy of Meta.
This article was generated with AI assistance and reviewed for accuracy and quality.
