Atlas

Models

← All models

DeepSeek-R1

DeepSeek · 2025-01-20 · 684.5B parameters

DeepSeek-R1 is DeepSeek's open-weight MoE reasoning checkpoint released January 20, 2025 and trained on top of DeepSeek-V3-Base. Its downloadable checkpoint contains exactly 684,489,845,504 model-weight parameters including one multi-token-prediction module; DeepSeek separately reports rounded figures of 671B main-model parameters, 37B activated per token, and 128K context support. Its multi-stage post-training combines cold-start supervised fine-tuning, reasoning-oriented GRPO, rejection-sampling supervised fine-tuning on about 600K reasoning and 200K non-reasoning examples, and a final reinforcement-learning stage.

Benchmark scores

BenchmarkScore
AA-LCR52.3
AA-Omniscience Index-31.3
Agent Red Teaming k=100 (CAIS)90.9
Agentic Tool Use (Chat)60.9
Agentic Tool Use (Enterprise)65.3
AidanBench1448
Aider Polyglot56.9
AIME 2024 (avg@8)77.5
AIME 2024-202574.0
AIME 202570.0
AIME 2025 (avg@8)70.4
ARC-AGI-115.8
ARC-AGI-21.3
Arena Text — Business, Management & Financial Ops (No Style Control)1347
Arena Text — Chinese (No Style Control)1400
Arena Text — Creative Writing (No Style Control)1355
Arena Text — English (No Style Control)1385
Arena Text — Entertainment, Sports & Media (No Style Control)1347
Arena Text — Exclude Ties (No Style Control)1346
Arena Text — Exclude Ties (Style Controlled)1383
Arena Text — Expert (No Style Control)1337
Arena Text — French (No Style Control)1366
Arena Text — French (Style Controlled)1390
Arena Text — German (No Style Control)1385
Arena Text — Hard Prompts (No Style Control)1361
Arena Text — Hard Prompts English (No Style Control)1376
Arena Text — Instruction Following (No Style Control)1358
Arena Text — Japanese (No Style Control)1325
Arena Text — Korean (No Style Control)1330
Arena Text — Legal & Government (No Style Control)1371
Loading Atlas data…