AI2 Olmo 3.1: Reinforcement Learning for Stronger Reasoning

Hear’s a summary of the key information from the provided text:

* New Models Released: The Allen Institute for AI (AI2) has released Olmo 3.1, an updated version of their Olmo 3 language models.
* improved Performance: Olmo 3.1 is described as a “much more performant sibling” to Olmo 3 Instruct 7B and is intended for real-world applications.
* Benchmark Results:

* Olmo 3.1 Think: Outperformed Qwen 3 32B on the AIME 2025 benchmark and performed similarly to Gemma 27B.
* Olmo 3.1 instruct: Performed well against other open-source models, even surpassing Gemma 3 on the Math benchmark.
* Availability: The models are currently available on the Ai2 Playground and Hugging Face. API access is coming soon.