Atlas

Models

← All models

DeepSeek-V3

DeepSeek · 2024-12-26 · 684.5B parameters

DeepSeek-V3 is DeepSeek's open-weight Mixture-of-Experts chat model, released December 26, 2024 and marketed as a 671B-parameter main model with 37B parameters activated per token; its released checkpoint also includes one multi-token-prediction module. The base model was pretrained on 14.8 trillion tokens, extended from 4K to 128K context with YaRN, and then aligned through supervised fine-tuning and GRPO reinforcement learning with reasoning patterns distilled from DeepSeek-R1.

Benchmark scores

BenchmarkScore
AA-LCR29.0
AA-Omniscience Index-41.4
Agentic Tool Use (Chat)58.5
Agentic Tool Use (Enterprise)62.5
AIME 2024 (avg@8)30.4
AIME 2024-202527.5
AIME 2025 (avg@8)24.6
ARC AI295.3
Arena Text — Business, Management & Financial Ops (No Style Control)1307
Arena Text — Chinese (No Style Control)1338
Arena Text — Creative Writing (No Style Control)1330
Arena Text — English (No Style Control)1346
Arena Text — Entertainment, Sports & Media (No Style Control)1324
Arena Text — Exclude Ties (No Style Control)1289
Arena Text — Exclude Ties (Style Controlled)1328
Arena Text — Expert (No Style Control)1305
Arena Text — French (No Style Control)1342
Arena Text — French (Style Controlled)1371
Arena Text — German (No Style Control)1324
Arena Text — Hard Prompts (No Style Control)1312
Arena Text — Hard Prompts English (No Style Control)1325
Arena Text — Instruction Following (No Style Control)1316
Arena Text — Japanese (No Style Control)1267
Arena Text — Korean (No Style Control)1249
Arena Text — Legal & Government (No Style Control)1341
Arena Text — Life, Physical & Social Science (No Style Control)1333
Arena Text — Longer Query (No Style Control)1343
Arena Text — Medicine & Healthcare (No Style Control)1345
Arena Text — Multi-turn (No Style Control)1349
Arena Text — Non-English (No Style Control)1315
Loading Atlas data…