Atlas

Models

← All models

DeepSeek-V3.2-Exp

DeepSeek · 2025-09-29 · 685.4B parameters

DeepSeek-V3.2-Exp is DeepSeek's experimental open-weight MoE checkpoint, released September 29, 2025 after continued training from DeepSeek-V3.1-Terminus to add DeepSeek Sparse Attention (DSA), a lightning-indexer-driven fine-grained top-k mechanism instantiated under MLA. Its continued pre-training used a dense indexer warm-up followed by sparse training of all parameters, and its post-training retained V3.1-Terminus's specialist-distillation and mixed-GRPO pipeline. It supports chat-template-selectable thinking and non-thinking modes with a documented 128K-token context window.

Benchmark scores

BenchmarkScore
A-Fantasia73.3
A-Fantasia Backwards Spelling93.0
A-Fantasia Chess59.0
A-Fantasia Cube Rotation68.0
AA-LCR69.0
AA-Omniscience Index-29.9
Aider Polyglot74.2
AIME 202557.7
Apex (MathArena)0.5
Arena Code WebDev — Brand & Marketing1271
Arena Code WebDev — Consumer Product1284
Arena Code WebDev — Content Creation Tools1243
Arena Code WebDev — Data & Analytics1260
Arena Code WebDev — Gaming1278
Arena Code WebDev — HTML1282
Arena Code WebDev — Reference-Based Design1276
Arena Code WebDev — Simulations1258
Arena Text — Business, Management & Financial Ops (No Style Control)1413
Arena Text — Chinese (No Style Control)1463
Arena Text — Creative Writing (No Style Control)1404
Arena Text — English (No Style Control)1441
Arena Text — Entertainment, Sports & Media (No Style Control)1405
Arena Text — Exclude Ties (No Style Control)1414
Arena Text — Exclude Ties (Style Controlled)1416
Arena Text — Expert (No Style Control)1420
Arena Text — French (No Style Control)1436
Arena Text — French (Style Controlled)1448
Arena Text — German (No Style Control)1434
Arena Text — Hard Prompts (No Style Control)1430
Arena Text — Hard Prompts English (No Style Control)1446
Loading Atlas data…