Meta's Muse Spark 1.3 Surpasses Top AI Models

Meta released Muse Spark 1.3 earlier this week. On the Artificial Analysis Intelligence Index, it sits among the leading models, level with Claude Fable 5 and just behind the top Claude and OpenAI entries.

The chart also shows the earlier models in the series. Muse Spark 1.2 was already a major advance on coding and agentic work. 1.3 continues that trajectory. Taken together, the Muse Spark releases represent a clear shift: Meta moved from lagging on frontier benchmarks to competing near the top in under five months.

This improvement can be attributed to the large investments in data and talent. Meta paid $14.3 billion for a 49% stake in Scale AI and installed founder Alexandr Wang as chief AI officer to lead Meta Superintelligence Labs. The data side of the effort was more controversial. Reports described the forced reassignment of thousands of Meta engineers (estimates put it at 30-50% of core product, infrastructure, and security teams, totalling around 6,500 people) into an Applied AI unit focused on generating training data and labelling work. Internal accounts called the unit “the gulag.” Engineers who had been building products were suddenly writing puzzles and coding problems full-time for model training.

Another program tracked employee keystrokes, mouse activity, and screen content to harvest additional training data. More than 1,600 staff signed a petition against it. Zuckerberg later acknowledged mistakes in how the restructuring was handled. The combination of the Scale AI deal, the large-scale internal reassignments, and the surveillance-style data collection formed the operational backdrop to the rapid model progress.

I guess all these acquisitions and controversial decisions finally paid off for Meta.