Ai2 brings three open models to COLM 2026: Olmo Hybrid, Asta, and Bolmo

Three projects and an open-science strategy

Ai2 arrived at COLM 2026 in San Francisco with a portfolio of open research reflecting the lab's strategy: build research artifacts that others can study and modify, and develop AI tools for science in partnership with domain experts. The conference hosted an Open Science, Open House session, where Asta product lead Rubén Lozano and Olmo scale lead Iz Beltagy discussed the philosophy behind open development with attendees. As a nonprofit lab, Ai2 pursues fundamental scientific questions while publishing models, data, code, and methodology. For the organization, open collaboration isn't a secondary consideration — it's the primary way the work gets done.

Olmo Hybrid: an architecture built for efficiency

The conference featured the paper "Olmo Hybrid: From Theory to Practice and Back," which examines an architecture combining transformer attention (which retrieves specific details from earlier text) with linear recurrent layers (which maintain a compact state). The research ties theoretical work on the representational capacity of these architectures to controlled training experiments.

Olmo Hybrid delivered a measurable gain: it matched the accuracy of the baseline Olmo 3 7B model on MMLU benchmarks (which cover knowledge across a wide range of subjects) while using 49% fewer training tokens. These results are feeding into the development of the next Olmo version, currently in pretraining with a hybrid mixture-of-experts architecture that combines attention and recurrence, routing each token through a subset of the model's components.

Ai2 has also opened up the infrastructure behind this work. Olmo-core 3, released ahead of COLM, includes a redesigned system for training large mixture-of-experts models. Researchers can use the code to develop their own models and study training decisions, alongside checkpoints and technical reports from the broader Olmo ecosystem. Google recently reproduced Olmo 3 7B's training run in MaxText on Google Cloud TPUs, underscoring the reproducibility of the fully open approach.

Asta: a platform for agentic science

At the conference, senior director of NLP research Noah A. Smith spoke with communications lead Kyle Wiggers about the next iteration of Olmo and ongoing work on agentic models. Scientific tools need to match the real tasks researchers actually face — which is why Ai2 develops Asta together with scientists, studying both the capabilities and the limitations of these systems.

Ai2 recently released AstaBrief, an open model that generates citable reports from scientific questions and retrieved literature. Researchers can use it through the Fast mode on the Asta platform or download the model to run on their own infrastructure. Ai2 works directly with scientific communities to identify where general-purpose models fall short and what capabilities future open tools need to support. Senior research scientist Bodhisattwa Prasad Majumder attended COLM to extend these conversations, presenting several related papers.

Bolmo: text through bytes

During the conference, Nature published "Retrofitting language models to operate over bytes," the research underlying Bolmo. Most language models split text into chunks from a fixed vocabulary. Bolmo instead operates on the raw bytes representing characters, capturing details that subword tokenization tends to obscure. The paper describes how an existing language model can be converted to this approach with relatively little additional training.

Alongside the publication, Ai2 released additional checkpoints that let researchers experiment with the architecture themselves. By sharing both the findings and the materials, Ai2 gives other scientists the ability to test the approach and push it further.