Meta Platforms Inc. introduced a new flagship large language model today to improve complex multi-agent automation workflows.
Muse Spark 1.1 debuts with context compaction
The model, named Muse Spark 1.1, is accessible through Meta’s AI chatbot service and a public preview of the Meta Model API. Developers can now embed it into custom software, enabling businesses to create applications where multiple AI agents collaborate.
These workflows usually include a primary agent that plans tasks and secondary agents that carry them out. Muse Spark 1.1 can modify plans mid-task if new information emerges, reducing the need for human oversight. It also tackles a frequent issue in AI systems: limited context windows.
When agents produce large amounts of data during multi-step tasks, some details are often lost if they exceed the model’s capacity. Muse Spark 1.1 compresses the data while keeping essential information intact, allowing it to reference earlier steps even after long sequences. The model supports a context window of 1 million tokens, a substantial expansion over many existing systems.
Meta evaluated the model by tasking it with building a chat application from prompts. The AI captured interface screenshots, detected technical problems, located the relevant code, and applied fixes without human input. In benchmark tests, Muse Spark 1.1 achieved a score of 72.2 on the Vibe Code Bench v1.1, surpassing Meta’s previous flagship model by over 50 points. It also performed 18% better on the SWE-Atlas Codebase QnA test.
These capabilities extend beyond coding. The model can create e-commerce listings from product videos, place restaurant orders, or manage other tasks requiring coordination between agents. Businesses may see faster automation of routine processes, though practical adoption will depend on how easily developers can tailor the model to their needs.
Related: Jamf debuts Beacon threat hunting service for Mac
Muse Spark 1.1’s launch coincides with Meta’s broader AI infrastructure expansion. The company plans to grow its data center capacity to 14 gigawatts next year, partly to support demand for its custom AI chips. The Iris processor, set for mass production in September, is likely the MTIA400—a chip Meta previewed earlier this year. It features 51% more high-bandwidth memory than its predecessor and supports improved data formats, resulting in a 400% speed increase for AI workloads.
Custom silicon could open new enterprise options
The Meta Model API, which provides access to Muse Spark 1.1, operates on Meta’s infrastructure. The company’s investment in custom silicon suggests it may eventually offer more than cloud-based services. One potential product is on-premises inference appliances combining MTIA chips with the new model, allowing enterprises to run AI workloads locally without third-party cloud providers.
Other tech companies already sell custom AI chips to data centers. Meta’s entry could help it compete in this space, though enterprise adoption of its hardware remains uncertain. For now, attention centers on the model’s performance and whether its multi-agent features meet the reliability standards businesses require.
The company did not reveal pricing for the Meta Model API, but the public preview indicates a tiered access model based on usage may follow. Real-world performance will determine its success, as unpredictable variables often disrupt even well-planned workflows.
Meta’s latest release reflects broader trends in AI development.
