Ai2 at COLM 2026: Open research, from models to agents for science

Ai2 attended COLM 2026 in San Francisco to share its latest work on open language models, training infrastructure, and AI for science. The lab presented the Olmo Hybrid model, which merges transformer attention with linear recurrent layers; the hybrid design achieved the same MMLU accuracy as Olmo 3 7B while using 49% fewer training tokens, illustrating improved training efficiency. Ai2 also released a redesigned system for training large mixture‑of‑experts models, providing code, checkpoints, and technical reports for researchers to build and investigate their own models. Google Cloud TPUs replicated Ai2’s Olmo 3 7B training run in MaxText, underscoring the reproducibility of the fully open approach. In addition, Ai2 introduced Bolmo, a byte‑level language model that operates on raw bytes rather than subword tokenization, and released checkpoints for community experimentation. The lab also discussed its agentic platform Asta, which generates cited reports from research questions and retrieved literature, and is available in a fast mode or for local deployment. Ai2’s senior director of NLP research, Noah A. Smith, and senior research scientist Bodhisattwa Prasad Majumder highlighted the importance of building scientific tools with researchers and studying their capabilities and limits. The conference emphasized Ai2’s commitment to open collaboration, reproducibility, and advancing AI tools that serve scientific communities.