Atlas

Models

← All models

QwQ-32B

Alibaba · 2025-03-06 · 32.8B parameters

QwQ-32B is Alibaba's 32,763,876,352-parameter open-weight reasoning checkpoint, released March 6, 2025 after the separate QwQ-32B-Preview. It was post-trained with supervised fine-tuning and multi-stage reinforcement learning for math, coding, and general capabilities. Its documented maximum context is 131,072 tokens, with YaRN required for prompts longer than 8,192 tokens.

Benchmark scores

BenchmarkScore
AA-LCR25.0
Aider Polyglot20.9
Arena Text — Business, Management & Financial Ops (No Style Control)1321
Arena Text — Chinese (No Style Control)1378
Arena Text — Creative Writing (No Style Control)1289
Arena Text — English (No Style Control)1349
Arena Text — Entertainment, Sports & Media (No Style Control)1280
Arena Text — Exclude Ties (No Style Control)1280
Arena Text — Exclude Ties (Style Controlled)1289
Arena Text — Expert (No Style Control)1325
Arena Text — French (No Style Control)1339
Arena Text — French (Style Controlled)1347
Arena Text — German (No Style Control)1312
Arena Text — Hard Prompts (No Style Control)1326
Arena Text — Hard Prompts English (No Style Control)1340
Arena Text — Instruction Following (No Style Control)1298
Arena Text — Japanese (No Style Control)1261
Arena Text — Korean (No Style Control)1279
Arena Text — Legal & Government (No Style Control)1337
Arena Text — Life, Physical & Social Science (No Style Control)1347
Arena Text — Longer Query (No Style Control)1309
Arena Text — Medicine & Healthcare (No Style Control)1341
Arena Text — Multi-turn (No Style Control)1314
Arena Text — Non-English (No Style Control)1307
Arena Text — Non-English (Style Controlled)1310
Arena Text — Occupational Mathematical (No Style Control)1356
Arena Text — Polish (No Style Control)1290
Arena Text — Polish (Style Controlled)1292
Arena Text — Russian (No Style Control)1294
Arena Text — Software & IT Services (No Style Control)1348
Loading Atlas data…