Meta has launched Muse Spark Open, an open model with 816 billion parameters, 42 billion of them active. Meta positions it as a response to the open-weight models coming out of China.
On the TheoPhysEval benchmark, GPT-6-Astra matched leading specialists on 49 of 50 unpublished theoretical physics calculations, and some apparent failures exposed errors in the experts' own answers. The researchers say physics has not yet seen a breakthrough like mathematics because the agents lack taste, but they think autonomous top-level AI research could come soon.
Nvidia has released new Nemotron models, including Nemotron 3.5 Super VL. They join a wave of open-weight releases from US labs that followed the Hugging Face hack.
Google's new model, available since 13 October, positions itself as a credible option for Google Cloud and Workspace customers. JDN put it through three real-world tests, costs included.
Version 1.5 of the Apertus model is out, with its release notes published on the Hugging Face blog.
Claude Opus 5.5 has replaced Fable 5.1 at the top of Arena's AI agent leaderboard. On Arena's new alignment index, however, OpenAI takes the top five spots.