Ai2 Brings Olmo Hybrid, Asta and Bolmo to COLM 2026: Open Models and Agents

Ai2 arrived at COLM 2026 with three names on its agenda: Olmo Hybrid, Asta and Bolmo. The institute frames its participation around two themes. The first is open language models and the infrastructure used to train them. The second is AI for science. The post itself contains no figures. It is a conference recap, not a release, so it gives neither model sizes nor benchmark results.

What exactly Ai2 showed

Olmo is Ai2's family of language models, which the institute releases as openly as possible: weights, data, code and training checkpoints. Judging by its name, Olmo Hybrid belongs to the hybrid architectures, but the announcement describes neither its design nor its scale, so we won't guess at the details. Specifics should be in the project's technical paper and repository.

Bolmo is a line of byte-level language models. These models work directly on the bytes of the text and do without a conventional tokenizer. That removes the dependence on a fixed vocabulary and, in theory, makes it easier to handle rare languages, typos and nonstandard text. The price is longer sequences and compute trade-offs.

Asta is Ai2's effort on agents for science. The idea is that AI helps a researcher find literature, work through papers and carry out routine research steps, with the quality of such systems checked on public tests.

Why COLM matters for this agenda

COLM is a conference devoted entirely to language modeling. The audience is not a marketing one: people read the code, look at reproducibility and ask uncomfortable questions about data. For Ai2 it is a natural venue. The institute builds its position on openness, and any claim along the lines of "our models can be verified" gets verified here, literally.

The pairing of models and agents is telling too. Ai2 works on two levels at once: base models and the training stack, and on top of them, applied systems for researchers. This is how an academic lab tries to keep pace with closed companies without giving up publishing its results.

What the announcement tells us

• Event: COLM 2026, where Ai2 participates as a research organization.
• Projects: Olmo Hybrid, Asta, Bolmo.
• Topics: open language models, training infrastructure, AI for science.
• Format: a recap post about engaging with the research community.
• Metrics, model sizes and benchmark results: not disclosed.

What to do with this information in practice

If you are a researcher or engineer, don't treat the post as a source of data. Open the Ai2 project pages, find the papers and repositories for Olmo Hybrid and Bolmo, and compare the licenses, data composition and claimed results against your own tasks.

Teams that work with multilingual or "dirty" text should test Bolmo's byte-level approach separately on their own data. You will have to measure any quality gain yourself, because the announcement offers no universal figures.

For those who work on scientific workflows, Asta is a reason to try agents on a narrow task, such as a literature review on a single topic. Start with scenarios you can verify, where the result can be checked by hand.