Ai2 recaps COLM 2026: Olmo Hybrid, next Olmo, AstaBrief and Bolmo
Ai2 shared a range of open research at COLM 2026 in San Francisco, spanning model architecture and agents for science. Its paper "Olmo Hybrid: From Theory to Practice and Back" reports that combining transformer attention with linear recurrent layers reached the same MMLU accuracy as Olmo 3 7B using 49% fewer training tokens, and that work is informing the next Olmo, now in pre-training with a hybrid mixture-of-experts architecture. Ai2 also released Olmo-core 3, a redesigned system for training large MoE models, described AstaBrief, an open-weights model that generates cited reports from research questions and retrieved literature, and said its Bolmo paper on retrofitting language models to operate over bytes was published in Nature during the conference, with additional checkpoints released.
As a nonprofit lab publishing models, data and code, Ai2 gives outside researchers material they can reproduce and extend in the pursuit of more efficient model training.