Atlas

Models

← All models

Olmo 3 32B Think

Allen Institute for AI (Ai2) · 2025-11-20 · 32.2B parameters

Olmo 3 32B Think is Ai2's fully open 32B reasoning checkpoint, built from Olmo 3 Base through thinking SFT, DPO, and 750 steps of reinforcement learning with verifiable rewards. It generates intermediate thinking traces for math, coding, and general problem solving, and Ai2 publishes the weights, data, code, and checkpoints across the model flow.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-58.4
AIME 202573.7
Arena Text — Business, Management & Financial Ops (No Style Control)1291
Arena Text — Chinese (No Style Control)1298
Arena Text — Creative Writing (No Style Control)1254
Arena Text — English (No Style Control)1343
Arena Text — Entertainment, Sports & Media (No Style Control)1257
Arena Text — Exclude Ties (No Style Control)1238
Arena Text — Exclude Ties (Style Controlled)1247
Arena Text — Expert (No Style Control)1287
Arena Text — German (No Style Control)1285
Arena Text — Hard Prompts (No Style Control)1303
Arena Text — Hard Prompts English (No Style Control)1334
Arena Text — Instruction Following (No Style Control)1276
Arena Text — Legal & Government (No Style Control)1296
Arena Text — Life, Physical & Social Science (No Style Control)1312
Arena Text — Longer Query (No Style Control)1295
Arena Text — Medicine & Healthcare (No Style Control)1313
Arena Text — Multi-turn (No Style Control)1290
Arena Text — Non-English (No Style Control)1252
Arena Text — Non-English (Style Controlled)1261
Arena Text — Occupational Mathematical (No Style Control)1331
Arena Text — Polish (No Style Control)1244
Arena Text — Polish (Style Controlled)1250
Arena Text — Russian (No Style Control)1256
Arena Text — Software & IT Services (No Style Control)1333
Arena Text — Writing, Literature & Language (No Style Control)1272
Arena Text Overall — 25% Factuality1320
Artificial Analysis Intelligence Index6.4
Loading Atlas data…