Meta launches flagship Muse Spark 1.1 model with multi-agent upgrades

Meta Platforms Inc. today launched a new flagship large language model optimized to power multi-agent automation workflows.

Muse Spark 1.1 is available in the company’s Meta AI chatbot service and via an application programming interface. The Meta Model API, as it’s aptly called, will enable developers to embed the LLM in their custom software. The service is in public preview.

Multi-agent automation workflows usually comprise a main agent that develops a plan for how to perform tasks and subagents that carry out its instructions. The main agent generates its plan at the start of a new project. According to Meta, Muse Spark 1.1 can detect and respond to mid-task developments that make it necessary to change the project plan.

Tasks that require multiple agents to automate usually comprise a large number of steps. As a result, agents generate a significant amount of data while performing those steps. If the dataset exceeds the underlying LLM’s context limit, some of the information has to be discarded, which often comes at the expense of output quality.

Muse Spark 1.1 addresses the challenge with a context compaction mechanism. According to Meta, it compresses the data generated by AI agents in a way that saves the most important details. That enables it to retrieve “information from much earlier work” when it needs to carry over data from one sub-task to another. The model has a context window of 1 million tokens.

The new model’s context compaction and multi-agent features make it adept at coding tasks. In one internal test, Meta engineers had it generate a chat app based on prompts. It took screenshots of the program’s interface, spotted several technical issues, identified the code snippets that caused the issues and fixed them.

Muse Spark 1.1 scored 72.2 on an AI programming benchmark called Vibe Code Bench v1.1, which puts it more than 50 points ahead of Meta’s previous flagship LLM. It achieved a nearly 18% higher score on a second test called SWE-Atlas Codebase QnA.

The features that make the model adept at generating code also enable it to tackle other use cases. It can generate e-commerce listings based on a product video, place restaurant orders on the user’s behalf, and perform other multistep tasks.

The Meta Model API through which developers can access Muse Spark 1.1 presumably runs on the Facebook parent’s own infrastructure. Reuters reported today that the company plans to boost its data center capacity to 14 gigawatts next year. The plan is said to revolve around Iris, an internally developed AI chip set to enter mass production in September.

The report indicates that Iris is likely the MTIA400, a custom AI processor Meta previewed in March. It includes 51% more high-bandwidth memory than the company’s previous-generation silicon. Additionally, the chip supports enhanced versions of MX8 and MX4, two data formats that AI models use to store their information. Meta says MTIA400 is 400% faster than its predecessor.

The company’s custom silicon gives it the option to follow up the Meta Model API with additional enterprise offerings. For example, the Facebook parent could theoretically sell on-premises inference appliances that combine MTIA chips with Muse Spark 1.1. Other hyperscalers are already moving to make their custom AI chips available to third-party data center operators.

Image: Meta

A message from John Furrier, co-founder of SiliconANGLE:

Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.

15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more

11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network.

About SiliconANGLE Media