Meta Details Muse Wonder's Active Memory, Active Focus, and Out-of-Distribution Scaling
Meta explains two features that make Muse Wonder suited for long-running agentic tasks—Active Memory and Active Focus—and reports that Muse Wonder Full outperfo…
Meta explains two features that make Muse Wonder suited for long-running agentic tasks—Active Memory and Active Focus—and reports that Muse Wonder Full outperfo…
Meta announced the Muse Wonder series, including Muse Wonder and Muse Wonder Full, frontier models designed for personal and professional agentic tasks, availab…
Mistral AI introduces Le Chonk, a visual grounding model that pinpoints object locations in images, scoring 42% on Dense 200 and targeting engineering, manufact…
Z.ai announced GLM-Legal, a legal-focused model built on GLM-5.3-Flash that it says matches frontier models on mainstream legal benchmarks and targets workflows…
A new atlas of microproteins in the human brain is available on Hugging Face, covering over 1,000 microproteins including one reduced in Alzheimer's that may af…
A call for researchers to openly share their RL environments, noting that one experiment's tooling can advance the whole field, especially amid reduced sharing…
An experiment tweaking the Pi harness for Qwen models found gains on ALE but little improvement on Terminal-Bench Lite, raising questions about why.
Requesty reports that Haiku 5.5 captured 51% of Haiku tokens in its first 46 hours and is 2.9x faster than Haiku 4.5, with 48% lower cost per request.
Command Code released an extension for VS Code and Cursor, expanding its product lineup to include CLI, Desktop App, and Extension, with web and mobile planned…
Hugging Face introduced a new 3D-Arena entry focused on LLM-generated 3D worlds, letting users compare and vote on outputs from multiple models.