Atlas

Models

← All models

Olmo 3 7B Instruct

Allen Institute for AI (Ai2) · 2025-11-20 · 7.3B parameters

Olmo 3 7B Instruct is Ai2's fully open 7B chat and quick-response checkpoint, produced from Olmo 3 Base through SFT, DPO, and reinforcement learning with verifiable rewards for multi-turn instruction following and tool use. Its base was explicitly extended from 8,192-token to 65,536-token sequences using YaRN on the full-attention layers, and Ai2 releases the weights, data, code, and intermediate checkpoints across the model flow.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-77.8
AIME 202541.3
Artificial Analysis Intelligence Index2.8
Artificial Analysis Omniscience Accuracy7.3
Artificial Analysis Omniscience Hallucination Rate91.8
Artificial Analysis Openness Index88.9
Capability116.9
CritPt0.0
GPQA Diamond40.0
Humanity's Last Exam (Text-Only)5.8
IFBench (Artificial Analysis)32.8
LiveCodeBench (Artificial Analysis)26.6
MMLU Pro52.2
SciCode (Artificial Analysis)10.3
SnakeBench Average Apples0.2
SnakeBench Best Apples1.0
SnakeBench Rating10.1
SnakeBench Total Apples4.0
SnakeBench Win Rate27.8
Terminal-Bench Hard0.0
τ²-bench (Telecom)12.6
Loading Atlas data…