Atlas

Models

← All models

OLMo 2 32B

Allen Institute for AI (Ai2) · 2025-03-13 · 32.2B parameters

OLMo 2 32B is Ai2's fully open dense base language model, pretrained on OLMo-Mix-1124 and mid-trained on Dolmino-Mix-1124; Ai2 describes its training exposure as up to 6 trillion tokens, and the released checkpoint is a merge of three 100B-token and one 300B-token mid-training runs. It was released alongside a separate OLMo 2 32B Instruct checkpoint derived with the Tülu 3.1 post-training recipe, with weights, data, code, logs, and intermediate checkpoints openly released.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AIME 20253.3
Artificial Analysis Intelligence Index5.0
Artificial Analysis Openness Index88.9
Capability110.6
GPQA Diamond32.8
Humanity's Last Exam (Text-Only)3.7
IFBench (Artificial Analysis)38.1
LiveCodeBench (Artificial Analysis)6.8
MMLU Pro51.1
SciCode (Artificial Analysis)8.0
Terminal-Bench Hard0.0
τ²-bench (Telecom)0.0
Loading Atlas data…