Summary:
1. The Allen Institute for AI (Ai2) has released the new Olmo 3.1 models, focusing on efficiency, transparency, and control for enterprises.
2. The updated Olmo 3.1 models show better performance on benchmarks, outperforming previous models like Olmo 3 and Qwen 3.
3. Ai2 remains committed to transparency and open source, allowing organizations to add to the model’s data mix and retrain it as needed.
Article:
The Allen Institute for AI (Ai2) recently unveiled its latest family of models, Olmo 3.1, designed to enhance efficiency, transparency, and control for enterprises. This new release builds upon the success of the previous Olmo 3 models, with updates to two key versions: Olmo 3.1 Think 32B and Olmo 3.1 Instruct 32B. These models have been optimized for advanced research, instruction-following, multi-turn dialogue, and tool use.
Ai2’s commitment to improving performance is evident in the benchmarks where the Olmo 3.1 models have excelled. Olmo 3.1 Think outperformed previous models like Qwen 3 32B in the AIME 2025 benchmark, showcasing substantial gains in math, reasoning, and instruction-following benchmarks. Similarly, Olmo 3.1 Instruct has proven to be a formidable contender, even surpassing models like Gemma 3 on the Math benchmark.
In addition to performance enhancements, Ai2 remains dedicated to transparency and open source practices. The Olmo 3.1 models allow organizations to contribute to the data mix and retrain the models as needed, fostering a collaborative and adaptive approach to AI development. This commitment to transparency is further reinforced by Ai2’s OlmoTrace tool, which tracks the alignment of LLM outputs with training data.
Overall, the release of Olmo 3.1 signifies Ai2’s continued efforts to advance AI capabilities while maintaining transparency and openness. By combining performance improvements with a commitment to collaborative development, Ai2 sets a new standard for AI innovation in the industry.
