Ai2 Newsletter

It's been a big month for fully open AI. We're excited to share two major releases: Olmo 3, our newest fully open language model family, and OlmoEarth, our Earth-AI model and platform. 

Olmo 3 is our new state-of-the-art family of open models that treats LMs not as frozen snapshots, but as a complete model flow, from pretraining through mid- and post-training. Instead of a single set of final weights, you get the full lifecycle: every major stage, checkpoint, dataset, and training decision needed to understand, adapt, and extend the models.

The Olmo 3 family includes:

Olmo 3-Think (7B, 32B)–our flagship open reasoning models for advanced experiments and RL, surfacing intermediate thinking steps.

Olmo 3-Instruct (7B)–tuned for multi-turn chat, tool use, and function/API calling.

Olmo 3-Base (7B, 32B)–strong at code, reading comprehension, and math; our best fully open base models and a versatile foundation for fine-tuning.

With a 65K-token context window, Olmo 3 handles full-paper Q&A, cross-document summaries, and deeper chain-of-thought reasoning in a single pass.

Alongside the models, we’re releasing datasets including Dolma 3 for pretraining, Dolmino for mid-training, and Dolci for post-training; industrial-grade data decontamination and deduplication tools; a full evaluation suite; and checkpoints at every major milestone, so you can customize and inspect models end-to-end.

You can try Olmo 3 in the Ai2 Playground, download the weights on Hugging Face, and explore the technical report and training recipes to plug the model flow directly into your own workflows.

OlmoEarth Platform is our end-to-end system that turns raw multi-sensor Earth data into decision-ready maps, alerts, and dashboards. 

Powered by our industry-leading OlmoEarth model family, the platform is already helping partners like Global Mangrove Watch, IFPRI, NASA JPL, Amazon Conservation Association, and the African Wildlife Foundation with tasks such as mangrove and habitat mapping, crop-type and cropland mapping, and forest and wildfire risk monitoring. 

Built with organizations across conservation, agriculture, and more, OlmoEarth is designed to adapt to new regions and use cases without requiring AI expertise—and while remaining open and free to use.

Explore here and here for additional details and how to get involved.

We also released Deep Research Tulu (DR Tulu), an open, end-to-end training recipe for long-form research agents—plus a DR Tulu-8B (RL) reference model you can use right away.

DR Tulu agents plan, search, and synthesize across web and scholarly sources to produce well-attributed answers with inline citations. They outperform prior open deep research systems on long-form benchmarks like ScholarQA-CSv2, ResearchQA, DeepResearchBench, and HealthBench while staying extremely affordable to run.

Read the DR Tulu blog for the full recipe, benchmarks, and links to models and code.

Thanks for being part of the Ai2 community. If you build on Olmo 3, OlmoEarth, or DR Tulu, we'd love to hear what you discover!