Atlas

Models

← All models

DeepSeek-V3.2

DeepSeek · 2025-12-01 · 685.4B parameters

DeepSeek-V3.2 is DeepSeek's open-weight MoE model, released December 1, 2025, and uses DeepSeek Sparse Attention: top-k sparse attention instantiated under MLA whose core attention is O(Lk), though its lightweight indexer remains O(L²). Its post-training combines specialist distillation with mixed GRPO-based reinforcement learning. It supports thinking and non-thinking modes, including tool use in both modes, with a documented 128K-token context window; its agentic task synthesis pipeline covered 1,800+ environments and 85,000+ prompts.

Benchmark scores

BenchmarkScore
AA-LCR65.0
AA-Omniscience Index-20.9
Agent Red Teaming k=100 (CAIS)93.0
AIME 2024 (avg@8)87.5
AIME 2024-202584.6
AIME 202559.0
AIME 2025 (avg@8)81.7
AIME 202694.2
Antidote: Everyday Edition954
Apex (MathArena)2.1
Apex Shortlist50.5
APEX-Agents7.0
APEX-Agents-AA14.5
APEX-SWE15.8
ARC-AGI-157.0
ARC-AGI-25.0
Arena Code WebDev — Brand & Marketing1366
Arena Code WebDev — Consumer Product1368
Arena Code WebDev — Content Creation Tools1364
Arena Code WebDev — Data & Analytics1381
Arena Code WebDev — Gaming1343
Arena Code WebDev — HTML1369
Arena Code WebDev — React1350
Arena Code WebDev — Reference-Based Design1352
Arena Code WebDev — Simulations1341
Arena Text — Business, Management & Financial Ops (No Style Control)1409
Arena Text — Chinese (No Style Control)1459
Arena Text — Creative Writing (No Style Control)1399
Arena Text — English (No Style Control)1441
Arena Text — Entertainment, Sports & Media (No Style Control)1395
Loading Atlas data…