Meta Launches Muse Wonder and Targets OpenAI

Meta has released a new series of models, Muse Wonder, featuring two variations and functions designed for multi-step tasks. The company claims that these systems have reached the level of players like OpenAI and Anthropic, a ramp-up that has already been internally referred to by the code name Watermelon.

Code Name Watermelon and Internal Power Boost

According to a source familiar with the matter, the new series was internally referred to as Watermelon. In July, it was reported that Meta's AI head, Alexandr Wang, informed staff that Watermelon had reached the level of OpenAI and emphasized that "Watermelon uses an order of magnitude of computation greater than Avocado." Avocado was the code name used for Muse Spark, a previous family of models. As the launch approached, Mark Zuckerberg and other employees shared posts themed around watermelon to signal the release.

Announced Capabilities: Multi-Step Tasks, Memory, and Exploration

Meta presents these models as capable of handling complex multi-step operations, such as disputing an insurance claim. The company highlights memory functions and the ability to try different approaches to accomplish these tasks.

Two Variants, Intended Uses, and Stated Positioning

The series includes two variations: Muse Wonder, a more compact and faster model that Meta plans to use for everyday tasks, and Wonder Full, a larger and more powerful version. Meta has introduced Muse Wonder and claims to have reached the level of labs such as Anthropic and OpenAI. Mark Zuckerberg presents Muse Wonder as a more significant and sophisticated category of models. This launch comes after the introduction of the Muse assistant and as the company invests hundreds of billions of dollars in AI, amounts that had raised concerns among some investors earlier this year.