Hugging Face Benchmark Leaderboard Analysis: Chinese Models Dominate, Zero Verifications

An analysis of Hugging Face public API data as of October 5, 2026, revealed that out of 48 datasets tagged with benchmark:official, 44 had leaderboards. A total of 1,023 submissions were confirmed on the default view, submitted by 411 models and 95 organizations.

Competitive Landscape and Regional Share

The median number of submissions per benchmark was 11.5, lower than the mean (21.3), indicating a right-skewed distribution where submissions are concentrated in a few benchmarks. The top 5 benchmarks (MMLU-Pro, GPQA, HLE, etc.) accounted for 45.9% of all submissions. By country, China dominated with 565 submissions (55.2%), followed by the US (22.6%) and South Korea (4.1%). Chinese organizations such as Qwen, Z.ai, and DeepSeek led in participation scale, but the number of #1 rankings was distributed among FINAL-Bench (9 times), Moonshot AI (4 times), and Z.ai (4 times).

Lack of Verification and Data Interpretation Caveats

The most significant finding is that all scores are self-reported values from creators. There were 0 items flagged as verified by Hugging Face. Additionally, the default leaderboard view excludes derived models with a declared base_model; including them would increase the submission count by approximately 30% (446 submissions). Agent and terminal-related benchmarks numbered 17, the most among categories, but due to evaluation costs, submissions per benchmark averaged only 11.6, indicating lower competitive intensity compared to science and knowledge domains (58.0 submissions per benchmark).