4 results for “pretraining”
NVIDIA released Nemotron 3.5 Lightning, a 30B MoE model with 3B active parameters built for high-volume agent execution. The open model cuts inference cost for always-on agents by handling routine tool calls while frontier models focus on planning.
Apple researchers introduced LINK, a pretraining data intervention that swaps English words with translations from bilingual vocabularies to improve knowledge transfer to low-resource languages. The method needs no extra training or parallel data and shows up to 2x faster convergence on downstream tasks across eight languages.
Series start (Aug 5, 2026): Kavukcuoglu named DeepMind SVP; Hassabis becomes chair and Alphabet chief scientist; Dean & Ghemawat exit to a PBC.
Liquid AI has released the LFM2.5 model family, a comprehensive suite of open-weight models designed to run entirely on edge devices without cloud. The LFM2.5 releases address three critical constraints that have limited on-device AI.