Olmo 3

What it is

Olmo 3 is the third generation of fully open language models from the Allen Institute for AI (Ai2). It was announced on 20 November 2025. The models are dense decoder-only Transformers at 7B and 32B parameters ^[1]. A Transformer is the neural network design that most modern language models use, and decoder-only means the model generates text one token at a time from the text before it.

Origins

Olmo 3 is the successor of OLMo 2 ^[1]. It continued the family with new base, reasoning, chat and reinforcement-learning variants, described below.

Main features

Olmo 3 comes in several variants. The Base models come in 7B and 32B sizes. The Think reasoning models also come in 7B and 32B sizes. The Instruct model, designed for chat and tool use, comes in a 7B size. RL Zero, a 7B model, is an open reinforcement-learning pathway that starts from the base model ^[1]. Reasoning models are trained to work through problems step by step before answering.

The pretraining used the Dolma 3 Mix, about 5.9 trillion tokens drawn from a Dolma 3 corpus of about 9.3 trillion tokens. This was followed by 100 billion tokens of Dolmino mid-training data and about 50 billion tokens of Longmino long-context data. The post-training data is called Dolci ^[1]. The context window is 65,536 tokens ^[2], which is the amount of text the model can consider at one time.

Ai2 called Olmo 3-Think 32B the strongest fully open thinking model. It scored 89.0 on IFEval, a test of instruction following, and was trained on roughly six times fewer tokens than Qwen 3 32B. Think 7B matched Qwen 3 8B on MATH, a benchmark of math problems ^[1]. Ai2 also called the 32B base model the strongest fully open base model ^[1]. These are Ai2's own claims.

Pretraining used up to 1,024 H100 GPUs. Mid-training used 128 and post-training used 256 ^[1]. The model launched alongside OlmoTrace in the Ai2 Playground. OlmoTrace traces a model's outputs back to its training data in real time ^[1].

The weights were released under the Apache 2.0 license, and all of the datasets were made openly available ^[1][2].

Olmo 3.1 was released on 12 December 2025. It extended the reinforcement-learning run for Think 32B, which gained more than 5 points on AIME and more than 20 points on IFBench ^[1][3]. AIME and IFBench are benchmarks used to test reasoning and instruction following. Olmo 3.1 also added an Instruct 32B model and refreshed the RL Zero 7B Code and Math checkpoints ^[1][3].

Place in the field

Olmo 3 is part of Ai2's line of fully open models, and it was followed by Olmo Hybrid 7B in March 2026, a model with a hybrid architecture ^[1]. Across the series, the models are released with their data, training details and tools, so that outside researchers can trace how each variant was made. Readers should note that the performance claims are Ai2's own reports.