AstaBrief, Olmo-core 3, and more from Ai2

This week, we’re releasing two new open artifacts for researchers and model builders: AstaBrief 8B, a model for generating fast, cited scientific reports, and Olmo-core 3, our redesigned framework for training large mixture-of-experts (MoE) models.

Together, they reflect our approach to truly open AI—sharing models and the infrastructure behind them so others can inspect how they work and adapt them to new problems. From scientific synthesis to large-scale model training, we want to make advanced AI systems more accessible to – and reproducible by – more people.

Heads up: We'll be at COLM 2026 in San Francisco, October 6–9. Read more about where to find us – and the work we'll be presenting – below.

AstaBrief

Reviewing and synthesizing scientific literature can be time-consuming. AstaBrief 8B is an 8-billion-parameter model trained to generate a cited report from a research question and retrieved scientific literature, giving researchers a faster starting point for exploring a topic, checking sources, and refining their questions.

AstaBrief was trained on scientific research queries and carefully filtered report examples designed to reinforce citation grounding. Unlike Asta’s Thinking mode, which uses proprietary models and multiple intermediate steps to build a report, AstaBrief generates the final synthesis in a single pass—helping make report generation much quicker and on hardware you control.

Across Asta’s full report-generation pipeline, Fast mode averages 51.1 seconds per report compared with 178.5 seconds for Thinking mode—about 3.5× faster.

We’re releasing AstaBrief's weights and training data alongside an example workflow for generating reports from your own PDFs—giving researchers an open resource to build and customize scientific synthesis tools.

Olmo-core 3

Training large MoE models can be more compute-efficient than dense models, but distributing their experts across GPUs creates new systems challenges as they scale.

Olmo-core 3 – the core infrastructure behind the next generation of Olmo – redesigns our open training stack around those challenges, making it easier to train large MoEs while preserving the efficiency that makes them attractive in the first place.

It does this in part by changing how experts are distributed during training. Instead of repeatedly gathering model weights for each small batch of data, Olmo-core 3 keeps experts resident on GPUs and routes data to them, reducing communication overhead as models grow.

These and other improvements let Olmo-core 3 scale to trillion-parameter configurations. See the experimental results in our tech report.

To make the underlying concepts easier to grasp, we've built an interactive experience that shows how MoE training scales from a single GPU to hundreds. Olmo-core 3 is now available in open source for researchers and developers to explore and build on.

Join our Discord | We're now on Reddit!

Goodfire traces unwanted model behavior with Olmo

Goodfire used Olmo checkpoints and our open training and evaluation tools to investigate unwanted changes in model behavior. The team traced a safety regression to individual training examples and tested targeted changes designed to reduce it while preserving broader capability gains.

Putting AutoDiscovery in students’ hands

University of Washington students used AutoDiscovery to explore open-ended scientific questions across materials science and biology. The challenge gave students experience interrogating AI-generated hypotheses while showing how tools like AutoDiscovery can surface promising directions for further investigation.

Steering Arena

Northeastern University student researcher Soham Padia turned an evaluation of Olmo 3’s prosocial behavior into a crowdsourced game. After roughly 600 submissions, participants found unexpected ways to optimize the game's internal signal and elicit certain responses from Olmo, showing how crowdsourced stress testing can uncover behaviors conventional evaluations miss.

Google reproduces Olmo 3 training on TPUs

Google’s MaxText team reproduced Olmo 3 7B’s pretraining and mid-training on Cloud TPUs, matching our original run on held-out evaluations. The work shows how a fully open model recipe can be independently implemented across different frameworks and hardware—and rigorously checked against the original.

Meet Ai2 at COLM 2026

We’re heading to COLM 2026 in San Francisco, October 6–9, with four days of workshops, posters, and talks on our latest AI research—from Olmo Hybrid to benchmarks for AI-assisted scientific writing.

Find us at booth 302 during the October 6–8 main conference to meet researchers behind our open models and AI for science work, including the teams behind AstaBrief and Olmo-core 3. And if you’re interested in building fully open AI with us, stop by to meet the team and learn about jobs at Ai2.

Work with us | Ai2 newsletter archive