Leading power, open roots: Olmo 3 introduces full model flow transparency and traceability, combining top-tier performance with increased efficiency and cost savings to advance open and sustainable AI.
SEATTLE–(BUSINESS WIRE)–The Allen Institute for AI (Ai2) today announced Olmo 3, a family of frontier fully open language models that delivers the entire model flow – the complete, transparent pipeline from data to deployment. The release includes the first-ever fully open 32B thinking model that generates explicit reasoning-chain-style content. Olmo 3 gives AI builders full visibility and control over every training stage, checkpoint, and dataset, enabling limitless customization and reproducible research at scale.
While AI capabilities have surged ahead this year, truly open models have lagged behind. Yet openness is essential to ensuring that progress remains transparent, explainable, and shared across the research community. Olmo 3 closes that gap, delivering cutting-edge reasoning and tool-use performance alongside complete traceability—including the ability to trace intermediate reasoning steps back to the training data and decisions that produced them. In a landscape where most models offer only limited inspectability and tooling, this level of openness is transformative. It cuts development time and resource demands, enabling faster, more resource-efficient iteration without starting from scratch.
Olmo 3 also shows that state-of-the-art performance can go hand in hand with lower energy use, reduced infrastructure costs, and a more sustainable path to industry-leading AI.
“Olmo 3 proves that openness and performance can advance together,” said Ali Farhadi, CEO of Ai2. “By sharing the full model flow, we’re giving researchers, developers, and builders everything they need for frontier-scale AI development that’s efficient, reproducible, and built to serve science. High performance doesn’t have to come at high cost, either. By dramatically reducing compute and energy requirements, we’re demonstrating that responsible, sustainable AI can scale without compromise.”
Olmo 3 empowers AI builders to inspect, modify, and reproduce the entire training process, unlocking infinite customization and unprecedented transparency. The Olmo 3 model flows range from 7B to 32B parameters, delivering high performance across devices – from laptops to high-end compute clusters:
- Olmo 3-Think (7B & 32B): Ai2’s latest flagship model. Olmo 3-Think 32B is a fully open, frontier-class reasoning model built for advanced research and RL experiments. Olmo 3-Think 7B brings reasoning transparency to a smaller scale, surfacing step-by-step thinking for complex tasks while retaining efficiency.
- Olmo 3-Base (7B & 32B): This model family delivers strong results in programming, reading comprehension, and math, maintains performance at extended context lengths, and integrates seamlessly with RL workflows – all trained on transparent data and open recipes. The base models are designed for continued pre-training and targeted fine-tuning and were used as the base for Olmo 3 Think and other post-trained models in the Olmo 3 family.
- Olmo 3-Instruct (7B): Another post-trained version of Olmo 3-Base, Olmo 3-Instruct is optimized for instruction following, multi-turn dialogue, and tool use.
All models are released under permissive open licenses (Apache 2.0), continuing Ai2’s leadership in truly open AI research.
“Transparency and performance are essential for developers to scale AI with open, U.S.-built models like Olmo 3,” said Kari Briski, vice president, Generative AI Software for Enterprise at NVIDIA. “Powered by NVIDIA AI infrastructure, these models turn intelligence into America’s most renewable resource, accelerating AI for researchers and industries.”
Built for Research. Ready for Use.
Olmo 3 isn’t just a model—it’s a complete model flow AI builder toolkit. Each model includes checkpoints from every major training phase and supports a 65K token context window – 16x larger than Olmo 2 – allowing reasoning across long documents and complex datasets.
The models are immediately available in Hugging Face and the Ai2 Playground.
Performance and Efficiency
The Olmo 3 family is the best collection of fully open base models, outperforming open models like Stanford’s Marin, LLM360’s K2, and Apertus. Models in the family benchmark on par with – or surpass – models many times their size across reasoning, comprehension, and long-context benchmarks.
Of note, Olmo 3-Think (32B) is the strongest fully open reasoning model, narrowing the gap to the best open-weight models of similar scale, such as the Qwen 3-32B-Thinking series of models across our suite of reasoning benchmarks, all while being trained on 6x fewer tokens. Olmo 3-Instruct outperforms open weight models, including Qwen 2.5, Gemma 3, and Llama 3.1, and is competitive with Qwen 3 model families of similar sizes (7B and 32B).
This efficiency means researchers and developers can achieve frontier-class results on far less compute, lowering both barriers to entry and environmental impact without compromising capability.
Transparent Data, Trustworthy AI
Olmo 3 is pretrained on Dolma 3, a 6-trillion-token open dataset spanning the web, scientific literature, and code – representing one of the most transparent and comprehensive corpora available for large-scale model training.
Every dataset and supporting tool, such as OlmoTrace for data traceability (will be enabled shortly after release), the new Dolci suite of datasets for reasoning fine-tuning, and OLMES for reproducible evals is fully open for inspection, replication, and reuse, ensuring end-to-end transparency and accountability.
Open for Collaboration
Olmo 3 represents a new era of open-first AI, where progress is shared, verifiable, and accessible to all. By opening every stage of development – from data to deployment – Olmo 3 empowers researchers and developers to trace model behavior back to its sources, understand how training choices shape outcomes, and build with confidence on a fully customizable foundation. Teams can fine-tune the models for new domains, experiment with alternative training objectives, or extend released checkpoints to drive fresh innovation across science, education, and real-world applications.
To learn more about Olmo 3, check out our blog post, read our technical report, explore the models and data, or try our new models in the Ai2 Playground.
About Ai2
Ai2 is a Seattle-based non-profit AI research institute with the mission of building breakthrough AI to solve the world’s biggest problems. Founded in 2014 by the late Paul G. Allen, Ai2 develops foundational AI research and innovative new applications that deliver real-world impact through large-scale open models, open data, robotics, conservation platforms, and more. Ai2 champions true openness through initiatives like OLMo, the world’s first truly open language model framework, Molmo, a family of open state-of-the-art multimodal AI models, and Tulu, the first application of fully open post-training recipes to the largest open-weight models. These solutions empower researchers, engineers, and tech leaders to participate in the creation of state-of-the-art AI and to directly benefit from the many ways it can advance critical fields like medicine, scientific research, climate science, and conservation efforts. For more information, visit allenai.org.
Contacts
