Meta Platforms through its dedicated artificial intelligence division Meta Superintelligence Labs launched Muse Spark 1.1 on July 9, 2026. This updated artificial intelligence software is a multimodal reasoning model built to handle complex autonomous jobs across multiple applications. The company also opened a public preview of the Meta Model API which lets outside software developers build applications using this architecture for the first time.
Key Takeaways
- Meta released Muse Spark 1.1 as a major update to its previous reasoning model.
- The model features a large context window of 1 million tokens and actively manages memory.
- Software developers can access the model via the new Meta Model API in public preview.
- The tool operates natively within multi-agent networks to run tasks in parallel.
- Everyday users can access the model for free in the Meta AI application by selecting Thinking mode.
Multi Agent Systems and Software Coding Gains
The new model coordinates multi-agent systems where one main software agent outlines a project plan and splits the workload among several parallel subagents. This structural design helps the system finish complex engineering operations faster by running tasks concurrently rather than processing them one after another. The model can work as either a primary supervisor or a targeted subagent that passes control back to the main manager when it encounters a roadblock.
For developers in India and worldwide who manage large software codebases, the model offers substantial upgrades in performance. Internal trials on the Meta Internal Coding Bench show that the system accurately identifies software bugs, introduces new features into large enterprise programs, and handles large code migrations. The model works with popular developer environments and supports planning features, goal conditions, and automated context reduction.
Desktop Automation and Context Window Management
The artificial intelligence system can manage a memory bank of 1 million tokens. It actively compacts conversation history, retains critical development steps, and pulls necessary data from earlier parts of a long working session. This process helps keep operational expenses lower because users do not have to resend the entire text history with every new prompt.
The software also introduces advanced computer control features. The system handles complex workflows across multiple separate software applications even when project requirements change during the job. The model evaluates whether to build automated scripts to save time or interact with the visual interface directly through clicks. Instead of executing one click at a time, the model builds a batch of multiple actions at each phase to complete tasks faster.
The model is now live for consumer testing on the meta.ai website and within the official Meta AI assistant app. For external developers using the new API platform, Meta set a pay-as-you-go pricing rate of 1.25 US dollars per 1 million input tokens and 4.25 US dollars per 1 million output tokens.
Frequently Asked Questions
Q1. What is Muse Spark 1.1?
A1. It is a multimodal reasoning artificial intelligence model built by Meta Superintelligence Labs that can process text, images, and audio to perform complex autonomous planning tasks.
Q2. How much does it cost developers to use the Meta Model API?
A2. The API uses a pay-as-you-go pricing structure that charges 1.25 US dollars per million input tokens and 4.25 US dollars per million output tokens.
Q3. Can general public users access the new model for free?
A3. Yes. General users can use the model for free by opening the Meta AI application or visiting the official website and switching the system into Thinking mode.
Q4. What is the size of the memory window in this version?
A4. The model features a context window of 1 million tokens and uses an active memory compression method to retain essential instructions during long chat sessions.
