Atlas

Models

← All models

DeepSeek-V3.2-Speciale

DeepSeek · 2025-12-01 · 685.4B parameters

DeepSeek-V3.2-Speciale is a high-compute experimental reasoning variant of DeepSeek-V3.2 released by DeepSeek on December 1, 2025. It was trained exclusively on reasoning data with a reduced length penalty during reinforcement learning and incorporates the DeepSeekMath-V2 dataset and reward method; unlike DeepSeek-V3.2, it does not support tool calling. DeepSeek reports gold-medal-level performance at IMO, CMO, IOI, and the ICPC World Finals 2025, with 2nd place at the ICPC World Finals and 10th place at IOI.

Benchmark scores

BenchmarkScore
AA-LCR59.3
AA-Omniscience Index-16.5
Apex (MathArena)9.4
Apex Shortlist69.7
Artificial Analysis Intelligence Index22.2
Artificial Analysis Omniscience Accuracy38.4
Artificial Analysis Omniscience Hallucination Rate88.9
BRUMO 202599.2
Capability149.4
CMIMC 202594.4
Creative Writing v3 Elo1221
Creative Writing v3 Rubric Score14.7
CritPt7.4
Global MMLU Lite88.4
GPQA Diamond87.1
HMMT Feb 202597.5
HMMT Nov 202593.3
Humanity's Last Exam (Text-Only)26.1
IFBench (Artificial Analysis)63.9
LisanBench Difficulty-Weighted1511
LisanBench Path Length5073
LiveCodeBench (Artificial Analysis)89.6
Longform Creative Writing48.8
MMLU Pro86.3
Pencil Puzzle Bench - Direct Ask2.3
Putnam 2025 (MathArena Accuracy)85.8
Ruby LLM Benchmarks — Program Fixer69.3
Ruby LLM Program Fixer — Calendar80.2
Ruby LLM Program Fixer — Parking Garage39.3
Ruby LLM Program Fixer — School Library80.1
Loading Atlas data…