Benchmark

DeepSeek — reported benchmarks

DeepSeek · source page ↗ · last checked Oct 19, 2026, 12:04 AM

Reported benchmarks · DeepSeek-V3.2

captured Oct 19, 2026, 12:04 AM
BenchmarkScore
MMLU-Pro85.0 % · Thinking mode
GPQA-Diamond82.4 % · Thinking mode
Humanity's Last Exam25.1 % · Thinking mode, text-only
LiveCodeBench83.3 pass@1 · COT, data from 2024.08-2025.04
Codeforces2386 Rating · Thinking mode
AIME 202593.1 % · Thinking mode
HMMT Feb 202592.5 % · Thinking mode
Terminal Bench 2.046.4 % · Code Agent, Thinking mode
SWE-bench Verified73.1 Resolved · Code Agent, Thinking mode
SWE-bench Multilingual70.2 Resolved · Code Agent, Thinking mode

Vendor-reported via automated web search — not independently verified. See the cited matrix on /models.

In the news · DeepSeek

Importance-filtered press coverage (Google News) mentioning DeepSeek. Headlines link to the original; verify before acting.

Change history