DeepSeek

DeepSeek-V3.2-Exp

Zero-eval
#1SimpleQA
#1MMLU-Pro
#1SWE-bench Multilingual
+2 more

by DeepSeek

+
+
+
+
About

DeepSeek-V3.2-Exp was introduced as an experimental release, designed to test new architectural innovations and training methodologies. Built to explore the boundaries of mixture-of-experts design, it serves as a research preview for techniques that may be incorporated into future stable releases.

+
+
+
+
Pricing Range
Input (per 1M)$0.27 -$0.27
Output (per 1M)$0.41 -$0.41
Providers2
+
+
+
+
Timeline
AnnouncedSep 29, 2025
ReleasedSep 29, 2025
+
+
+
+
License & Family
License
MIT
Performance Overview
Performance metrics and category breakdown

Overall Performance

14 benchmarks
Average Score
66.1%
Best Score
97.1%
High Performers (80%+)
4

Performance Metrics

Max Context Window
229.4K
Avg Throughput
100.0 tok/s
Avg Latency
1ms
+
+
+
+
All Benchmark Results for DeepSeek-V3.2-Exp
Complete list of benchmark scores with detailed information
SimpleQA
text
0.97
97.1%
Self-reported
AIME 2025
text
0.89
89.3%
Self-reported
MMLU-Pro
text
0.85
85.0%
Self-reported
HMMT 2025
text
0.84
83.6%
Self-reported
GPQA
text
0.80
79.9%
Self-reported
Aider-Polyglot
text
0.74
74.5%
Self-reported
LiveCodeBench
text
0.74
74.1%
Self-reported
CodeForces
text
0.71
70.7%
Self-reported
SWE-Bench Verified
text
0.68
67.8%
Self-reported
SWE-bench Multilingual
text
0.58
57.9%
Self-reported
Showing 1 to 10 of 14 benchmarks
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+