Atlas

Models

← All models

Qwen3-235B-A22B

Alibaba · 2025-04-29 · 235.1B parameters

Qwen3-235B-A22B is Alibaba's flagship open-weight MoE model in the Qwen3 family, released April 29, 2025. It has approximately 235B total parameters and approximately 22B activated parameters per token, 94 transformer layers, 128 total experts, and 8 activated experts per token. It supports both thinking and non-thinking modes, with a native 32,768-token context window validated up to 131,072 tokens using YaRN scaling.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-44.6
Aider Polyglot59.6
AIME 2024 (avg@8)84.6
AIME 2024-202584.0
AIME 202523.7
AIME 2025 (avg@8)83.3
Arena Text — Business, Management & Financial Ops (No Style Control)1391
Arena Text — Chinese (No Style Control)1426
Arena Text — Creative Writing (No Style Control)1358
Arena Text — English (No Style Control)1396
Arena Text — Entertainment, Sports & Media (No Style Control)1350
Arena Text — Exclude Ties (No Style Control)1371
Arena Text — Exclude Ties (Style Controlled)1386
Arena Text — Expert (No Style Control)1368
Arena Text — French (No Style Control)1371
Arena Text — French (Style Controlled)1385
Arena Text — German (No Style Control)1385
Arena Text — Hard Prompts (No Style Control)1394
Arena Text — Hard Prompts English (No Style Control)1392
Arena Text — Instruction Following (No Style Control)1361
Arena Text — Japanese (No Style Control)1369
Arena Text — Korean (No Style Control)1360
Arena Text — Legal & Government (No Style Control)1394
Arena Text — Life, Physical & Social Science (No Style Control)1406
Arena Text — Longer Query (No Style Control)1390
Arena Text — Medicine & Healthcare (No Style Control)1408
Arena Text — Multi-turn (No Style Control)1399
Arena Text — Non-English (No Style Control)1386
Arena Text — Non-English (Style Controlled)1393
Loading Atlas data…