Meta closes its AI just as China opens it

For years I've been hearing that Meta was the champion of open source in artificial intelligence: that Llama was Zuckerberg's gift to the community, the free counterweight to the closed labs. And then along comes Muse Spark. VentureBeat covered it with the headline "Goodbye, Llama?": Meta's first model since Superintelligence Labs was set up, and it's proprietary, with no published weights. The New Stack put it even more bluntly: Meta abandons Llama's open source for a closed model.

The detail that doesn't fit is on the other side of the board. While Meta closes the door, the Chinese labs throw it wide open: Qwen, DeepSeek, Kimi, GLM keep releasing weights under Apache 2.0 or MIT licenses nonstop, and Alibaba's Qwen family had already overtaken Llama in cumulative downloads. The one who preached openness shuts down; the one supposedly closed by geopolitical decree opens up.

I think that Meta's openness was never a conviction, it was a strategy to keep from being left out while it wasn't leading. Its half step back in August doesn't disprove it: it released Muse Glimmer, a 30-billion-parameter model with open weights under Apache 2.0, and promised to open Muse Spark 1.2; but the latest version, Muse Spark 1.3, is still closed and offered only through an API. The day it believes it has something to sell, it shuts the door, and nobody should be surprised. The uncomfortable part is the mirror: it turns out the one keeping the weights open is China, while the "free world" keeps them under lock and key. Maybe this "free" business was worth looking at piece by piece, not by flag.

Sources: VentureBeat · The New Stack · Artificial Intelligence News · Constellation Research · TrueFoundry (Muse Spark 1.3)