Know What's Happening in AI

Ai2's New Olmo 3.1 Extends Reinforcement Learning Training for Stronger Reasoning Benchmarks

The Allen Institute for AI unveiled Olmo 3.1, an upgraded 32‑billion‑parameter family that extends the original Olmo 3 models through an additional 21‑day reinforcement‑learning run on 224 GPUs. The Think 32B variant shows 5‑plus point gains on the AIME math benchmark and strong lifts on ZebraLogic, IFEval and IFBench, while the Instruct 32B model is tuned for chat, tool use and multi‑turn dialogue. Both models surpass Qwen 3 32B and rival larger open‑source contenders such as Gemma 27B. Checkpoints are now on the Ai2 Playground and Hugging Face, with API access slated for later this year.