Atlas

Models

← All models

DeepSeek-V3.1

DeepSeek · 2025-08-21 · 684.5B parameters

DeepSeek-V3.1 is DeepSeek's open-weight MoE model (marketed as 671B total with 37B active parameters) released August 21, 2025, introducing a hybrid reasoning architecture that switches between a fast Non-Thinking mode and a chain-of-thought Thinking mode via chat template. It also adds enhanced tool-use and agent capabilities with support for 128K-token context, under the MIT license.

Benchmark scores

BenchmarkScore
A-Fantasia75.3
A-Fantasia Backwards Spelling92.0
A-Fantasia Chess63.0
A-Fantasia Cube Rotation71.0
AA-LCR53.3
AA-Omniscience Index-28.4
AIME 202549.7
Apex (MathArena)0.5
Arena Text — Business, Management & Financial Ops (No Style Control)1409
Arena Text — Chinese (No Style Control)1470
Arena Text — Creative Writing (No Style Control)1403
Arena Text — English (No Style Control)1432
Arena Text — Entertainment, Sports & Media (No Style Control)1396
Arena Text — Exclude Ties (No Style Control)1406
Arena Text — Exclude Ties (Style Controlled)1404
Arena Text — Expert (No Style Control)1403
Arena Text — French (No Style Control)1457
Arena Text — French (Style Controlled)1453
Arena Text — German (No Style Control)1410
Arena Text — Hard Prompts (No Style Control)1417
Arena Text — Hard Prompts English (No Style Control)1424
Arena Text — Instruction Following (No Style Control)1403
Arena Text — Japanese (No Style Control)1368
Arena Text — Korean (No Style Control)1337
Arena Text — Legal & Government (No Style Control)1418
Arena Text — Life, Physical & Social Science (No Style Control)1433
Arena Text — Longer Query (No Style Control)1422
Arena Text — Medicine & Healthcare (No Style Control)1442
Arena Text — Multi-turn (No Style Control)1407
Arena Text — Non-English (No Style Control)1402
Loading Atlas data…