← Back to feed

Meta’s Superintelligence Lab released two new models on Monday: Muse Wonder and Muse Wonder Full. The company describes them as its first attempt at a frontier-class model, positioning them against OpenAI and Anthropic at the top of the capability stack.

The models are the second product series released under Alexandr Wang, who joined Meta as chief AI officer in June 2025 after founding Scale AI. The first series, Muse Spark, was a lightweight model designed for low-latency applications. Muse Wonder is a different category: larger, slower, and built to compete on capability rather than cost or speed.

The Watermelon Project

Internally, the project was code-named Watermelon. Bloomberg reported that Meta described the release as “a big leap” over its previous work, a claim that is typical of frontier model launches and will require independent benchmark evaluation to verify.

The release marks a sharp pivot for Meta’s AI strategy. For most of 2024 and 2025, the company’s public-facing AI efforts centered on the open Llama family, with commercial AI products running models from OpenAI and Anthropic. Wang’s mandate, as described by Meta when he joined, was to build a model that could anchor Meta’s own frontier position rather than relying on competitors’ infrastructure.

Muse Wonder Full is the larger of the two models. Based on available descriptions, Muse Spark aimed for lightweight performance; Wonder aims at raw capability. Specific benchmark numbers — SWE-bench, Arena ELO, LiveBench — were not available at time of writing. Stack Futures will update when independent evaluations are published.

Context: Meta’s AI Leadership Restructuring

Wang is a 2025 hire at a company that has historically struggled to retain senior AI talent at the research frontier. His arrival followed a period when Meta’s AI lab had lost several prominent researchers to OpenAI and other labs. The Superintelligence Lab was created as a distinct unit from Meta AI to give the frontier model effort autonomy and a separate resource allocation.

Muse Spark 1.3, released in September 2026, scored 81.6 overall on LiveBench in its xHigh effort configuration and was added to Arena’s Vision leaderboard in October. Muse Wonder Full is a substantially larger model. The Spark-to-Wonder gap mirrors the structure OpenAI uses between its o-series lightweight models and full frontier variants.

What Comes Next

Independent benchmark evaluation is the key unknown. If Muse Wonder reaches SWE-bench Verified scores above 85% or Arena ELO above 1480, it enters the cluster of models where capability differences are meaningful for enterprise buyers. Below those thresholds, the model competes primarily on price and Meta’s distribution scale — a different value proposition.

Meta’s distribution includes WhatsApp, Instagram, and Facebook, giving it direct consumer reach that OpenAI and Anthropic lack. If Muse Wonder is the engine behind Meta AI products, the deployment numbers will dwarf any lab’s API traffic regardless of where the model ranks on external benchmarks.