Atlas

Models

← All models

DeepSeek-V3.1-Terminus

DeepSeek · 2025-09-22 · 684.5B parameters

DeepSeek-V3.1-Terminus is an open-weight update to DeepSeek-V3.1, released September 22, 2025, that maintains the earlier model's capabilities while improving output stability. DeepSeek reports fewer Chinese-English mix-ups and abnormal characters, together with further optimization of Code Agent and Search Agent performance. It retains V3.1's chat-template-selectable thinking and non-thinking modes and documented 128K-token context.

Benchmark scores

BenchmarkScore
AA-LCR65.0
AA-Omniscience Index-24.3
AIME 202589.7
ALE-Bench1078
Arena Text — Business, Management & Financial Ops (No Style Control)1413
Arena Text — Creative Writing (No Style Control)1391
Arena Text — English (No Style Control)1431
Arena Text — Entertainment, Sports & Media (No Style Control)1389
Arena Text — Exclude Ties (No Style Control)1407
Arena Text — Exclude Ties (Style Controlled)1403
Arena Text — Hard Prompts (No Style Control)1427
Arena Text — Hard Prompts English (No Style Control)1433
Arena Text — Instruction Following (No Style Control)1406
Arena Text — Legal & Government (No Style Control)1457
Arena Text — Life, Physical & Social Science (No Style Control)1451
Arena Text — Longer Query (No Style Control)1420
Arena Text — Medicine & Healthcare (No Style Control)1467
Arena Text — Multi-turn (No Style Control)1415
Arena Text — Non-English (No Style Control)1405
Arena Text — Non-English (Style Controlled)1402
Arena Text — Polish (No Style Control)1427
Arena Text — Polish (Style Controlled)1417
Arena Text — Russian (No Style Control)1432
Arena Text — Software & IT Services (No Style Control)1427
Arena Text — Writing, Literature & Language (No Style Control)1410
Artificial Analysis Agentic Index18.1
Artificial Analysis Coding Index43.5
Artificial Analysis Intelligence Index30.4
Artificial Analysis Omniscience Accuracy27.9
Artificial Analysis Omniscience Hallucination Rate72.5
Loading Atlas data…