DeepSeek
deepseek-v3.2
mit
Benchmark Average (0–100)
68.2%
Average of non-null benchmark scores across all evaluated tasks.
BENCHMARK SCORES
GPQA82.4%
Graduate-Level Google-Proof Q&A. PhD-level multiple-choice questions in chemistry, biology, and physics. Scored by accuracy.
AIME 202593.1%
AIME 2025 is a mathematics competition benchmark testing advanced problem-solving.
SWE-bench Verified73.1%
SWE-bench Verified evaluates models on real-world software engineering tasks from GitHub issues.
BrowseComp51.4%
BrowseComp tests web browsing comprehension, measuring ability to find and synthesize information from websites.
HLE40.8%
Humanity's Last Exam (HLE) is a multi-modal benchmark testing frontier knowledge across mathematics, humanities, and natural sciences with 2,500 expert-level questions.
OTHER BENCHMARKS
ToolAthalon
35.2%
MODEL INFO
Organization
DeepSeek
Context Window
—
Release Date
2025-12-01
Knowledge Cutoff
—
Parameters
685000000000.0B
License
mit
METADATA
Announcement Date
2025-12-01
Organization Country
CN
Canonical ID
—