Ai2 invites ongoing discussion and collaboration by offering open papers, tools, and roles to contribute to ongoing research.

Ai2 emphasizes open collaboration as a core principle, openly sharing models, data, code, and methods to enable others to study and extend their work.

Olmo Hybrid combines transformer attention with linear recurrent layers, achieving training efficiency gains and equivalent accuracy while using 49% fewer training tokens on the MMLU benchmark.

Bolmo describes a Nature publication on retrofitting language models to operate over bytes, enabling handling of details hidden by subword tokenization, along with additional checkpoints for experimentation.

Google reproduced the Olmo 3 7B training run in MaxText on Google Cloud TPUs, underscoring reproducibility of Ai2’s open approach.

Ai2 discussed the next iteration of Olmo and agentic models in conversations with COLM attendees, including Noah A. Smith and Kyle Wiggers.

Asta is Ai2’s agentic platform for scientific work, featuring the open-weights AstaBrief model that generates cited reports from questions and literature, available in Fast mode or for local use.

The next Olmo model is in pre-training with a hybrid mixture-of-experts architecture, integrating attention and recurrence, and Olmo-core 3 provides a redesigned system for training large mixture-of-experts models to enable community experimentation.

Ai2 is actively engaging with scientific communities to identify where general-purpose models fall short and what future open models should support.

Ai2 attended COLM 2026 in San Francisco to share open research artifacts and collaborate with domain experts and the open-source community.