Atlas

Models

← All models

DeepSeek-V3-0324

DeepSeek · 2025-03-25 · 684.5B parameters

DeepSeek-V3-0324 is an updated checkpoint of DeepSeek-V3 (marketed as 671B total with 37B active parameters) released March 25, 2025, with an improved post-training pipeline borrowing RL techniques from DeepSeek-R1. It delivers substantially stronger reasoning and coding performance over the original V3 and was re-licensed under MIT.

Benchmark scores

BenchmarkScore
A-Fantasia68.7
A-Fantasia Backwards Spelling90.0
A-Fantasia Chess46.0
A-Fantasia Cube Rotation70.0
AA-LCR41.0
AA-Omniscience Index-37.6
Agentic Tool Use (Chat)54.3
Agentic Tool Use (Enterprise)64.2
AidanBench892
Aider Polyglot55.1
AIME 2024 (avg@8)58.6
AIME 2024-202552.2
AIME 2025 (avg@8)45.8
Arena Text — Business, Management & Financial Ops (No Style Control)1350
Arena Text — Chinese (No Style Control)1388
Arena Text — Creative Writing (No Style Control)1365
Arena Text — English (No Style Control)1386
Arena Text — Entertainment, Sports & Media (No Style Control)1357
Arena Text — Exclude Ties (No Style Control)1346
Arena Text — Exclude Ties (Style Controlled)1376
Arena Text — Expert (No Style Control)1351
Arena Text — French (No Style Control)1389
Arena Text — French (Style Controlled)1420
Arena Text — German (No Style Control)1380
Arena Text — Hard Prompts (No Style Control)1366
Arena Text — Hard Prompts English (No Style Control)1376
Arena Text — Instruction Following (No Style Control)1346
Arena Text — Japanese (No Style Control)1327
Arena Text — Korean (No Style Control)1317
Arena Text — Legal & Government (No Style Control)1380
Loading Atlas data…