The Allen Institute for AI (Ai2), a Seattle-based nonprofit research institute, has released Olmo 3, a new family of open-source language models. According to Ai2, the new models are designed to be competitive with other leading open models in performance and efficiency while offering a new level of transparency for developers and researchers.
The key feature of the Olmo 3 release is what Ai2 calls the “model flow.” Instead of just providing the final, trained model, the institute is making the entire development process public. This includes all training data, code, and intermediate versions of the model, known as checkpoints, from every stage of its creation. Ai2 states this approach allows users to understand how the model’s capabilities were developed, modify it for specific needs, and build greater trust in the system. An accompanying tool, OlmoTrace, lets users trace a model’s outputs back to the specific training data that influenced them.
The Olmo 3 family includes several versions tailored for different tasks:
- Olmo 3-Base: The foundational model, designed for tasks like programming, math, and reading comprehension.
- Olmo 3-Think: A “thinking” model that shows its step-by-step reasoning process, intended for complex problem-solving.
- Olmo 3-Instruct: A version optimized for conversational AI, following instructions, and using tools.
Ai2 claims that the Olmo 3 models achieve strong performance when compared to other open models of a similar size, such as Meta’s Llama 3.1, Google’s Gemma 3, and Qwen models. The institute also highlights the models’ training efficiency, stating that Olmo 3 was trained on significantly less data than some rivals, which can reduce computational costs and energy consumption.
The release is positioned to meet a growing demand from businesses and researchers for more customizable and transparent AI systems. By providing full access to the development process, users can adapt the models with their own proprietary data to create specialized applications. According to Ai2, this open approach gives organizations more control and confidence in the technology, ensuring they know exactly what data was used for training.
Olmo 3 is available under the permissive Apache 2.0 license, allowing for both commercial and research use. The models and their development components can be accessed through platforms like Hugging Face and the Ai2 Playground.
Sources: Allen AI Blog, GeekWire, VentureBeat
