Ai2 at COLM 2026: Open research, from models to agents for science
Allen Institute (AI2)
Ai2 showed new open research at COLM 2026, from a hybrid model to science tools. Its bet is simple: open code and checkpoints make AI easier to test, reuse, and trust.
Based on reporting by Allen Institute (AI2) — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ai2 spent this week at COLM 2026 in San Francisco, showing off work that fits its usual line: build AI research artifacts people can study and extend, then do the same for tools aimed at science. The lab also held an evening gathering that pulled in people from the COLM crowd and the wider open-source world.
That open setup is the point. Ai2 is a nonprofit lab, so it says it can chase basic research questions while sharing the models, data, code, and methods behind them. At COLM, that meant meeting the people who use those releases, argue with them, and sometimes improve them. The company frames that exchange as part of the research itself, not a nice extra.
One of the centerpieces was Olmo Hybrid: From Theory to Practice and Back. The paper looks at a language model that mixes transformer attention with linear recurrent layers. Attention can pull specific facts from earlier text, while recurrence keeps a compact running state as tokens move through the model. In Ai2’s experiments, the hybrid approach reached the same accuracy as Olmo 3 7B on MMLU while using 49% fewer training tokens.
That result is now feeding into the next Olmo model, which is in pre-training with a hybrid mixture-of-experts design. Ai2 also released Olmo-core 3 before the conference, adding a redesigned system for training large mixture-of-experts models. The code, checkpoints, and technical reports are all part of the broader Olmo ecosystem, and Google has already said it recreated the Olmo 3 7B training run in MaxText on Google Cloud TPUs.
Ai2 is pushing the same open approach into science work through Asta, its agentic platform for scientific tasks. AstaBrief, the newly released open-weights model, writes cited reports from research questions and retrieved literature. Scientists can use it in Asta’s Fast mode or run it on their own infrastructure, while Ai2 keeps working with scientific communities on where general-purpose models still come up short.
The conference also lined up with another release: Bolmo, the work behind “Retrofitting language models to operate over bytes,” which appeared in Nature. Instead of chopping text into pieces from a fixed vocabulary, Bolmo works on the bytes underneath the characters. Ai2 says that lets the model catch details that subword tokenization can miss, and the team released extra checkpoints so others can poke at the setup themselves.
My take — AI-written commentary, not fact-checked reporting
This is the rare open-AI pitch that sounds like engineering, not a sermon. Releasing the model, the code, the checkpoints, and the training system is harder than posting a glossy demo, which is exactly why it matters. Closed labs can keep their fireworks; the useful work still happens where other people can inspect the wiring.
Read more about this at: Allen Institute (AI2)