VIDRAFT Tops 10 Hugging Face Benchmarks as AI Leaderboard Map Goes Live

A new real-time dashboard reveals who is winning the global AI benchmarking race — and a Korean startup leads the pack.

TL;DR: Korean AI startup VIDRAFT holds first place on more Hugging Face official benchmarks than any other organization in the world, topping 10 out of 48 leaderboards as of October 5, 2026. QuantID, a Korean AI analytics firm, made this finding public by launching a real-time interactive map that aggregates all 48 Hugging Face official benchmark leaderboards into a single view. The map, built on live Hugging Face public APIs, also reveals that Chinese organizations dominate overall participation while competition in science and knowledge tasks is the most intense.

VIDRAFT, the Korean Pre-AGI AI startup, has claimed more first-place positions on Hugging Face's official benchmark leaderboards than any other AI organization globally, according to data published on October 6, 2026, by Korean tech outlet 디지털타임스. The revelation came via a newly launched real-time mapping tool built by QuantID (CEO Han Junhee) that consolidates all 48 Hugging Face official benchmarks into a single, scannable dashboard.

What VIDRAFT Announced

QuantID's new tool — titled Hugging Face Official Benchmarks: Map and Participation — pulls live data directly from Hugging Face's public API every time a user loads the page, meaning rankings and participation figures update in real time without QuantID storing or modifying the underlying data. The methodology, category classifications, and country attribution criteria are all disclosed publicly on the page itself.

Within this landscape, VIDRAFT (operating under the Hugging Face organization name FINAL-Bench) stands out as the single most successful organization by number of first-place holdings. As of October 5, 2026, VIDRAFT holds the top position on 10 benchmarks: GPQA Diamond, MMLU-Pro, MMMU-Pro, AIME 2026, HMMT 2026, LEXam, LEXam-hard, ExtractBench, IFStruct, and MDPBench. No other organization comes close — China's Zhipu AI holds 4 first-place spots, while Xiaomi, DeepSeek, and Moonshot AI each hold 3.

The dashboard itself surfaces several structural patterns across the global AI benchmarking ecosystem:

• Agent tasks dominate in count. Of the 48 official benchmarks, 17 focus on AI agent and terminal tasks — the largest single category. Science and knowledge benchmarks (6), document and OCR tasks (5), and coding, video, and legal benchmarks (4 each) follow.
• Science attracts the most submissions. Despite agents having the highest benchmark count, actual participation skews toward science and knowledge tasks, which account for 348 of approximately 1,023 total submissions. Agent benchmarks follow with 197 entries, and coding tasks with 158.
• Participation is highly concentrated. The median number of submissions per benchmark sits at around 12, yet the five busiest leaderboards — MMLU-Pro, GPQA, HLE, SWE-bench Verified, and Open ASR — account for 46% of all entries combined.
• China leads participation by country. Among 95 participating organizations, 55% are classified as Chinese, 23% as American, and 4% as Korean. The most active organizations by submission volume are Alibaba's Qwen team (136 submissions), Zhipu AI (84), DeepSeek and Moonshot AI (65 each), and Nvidia (63).
• Most organizations dip in, not dive in. Forty-five of the 95 participating organizations (47%) have submitted to only a single benchmark. By contrast, Alibaba's Qwen appears on 31 different leaderboards and Zhipu AI on 32.
• Top spots are razor-thin. Across the 42 leaderboards where a second-place entry exists, the median gap between first and second place is just 5.2 percentage points. Seven leaderboards show a margin of under 1%, meaning a single new submission could reshuffle rankings.
• Derivative models are largely hidden. Hugging Face's default leaderboard views exclude models built on top of other models. When those derivative submissions are counted, total entries rise to 1,469 — meaning roughly 30% of all submissions are invisible in the standard interface.
• Only 23 organizations hold any first-place title. VIDRAFT leads this exclusive group with 10, reinforcing its standing as the benchmark performance leader by a substantial margin.

Why It Matters

For enterprises and researchers evaluating AI models for deployment, the Hugging Face leaderboard ecosystem has long been a fragmented reference point — 48 separate pages with no unified view of relative strengths, competitive density, or geographic trends. QuantID's map addresses this usability gap directly, and in doing so has inadvertently spotlighted VIDRAFT's dominant technical performance across a remarkably diverse range of tasks: from doctoral-level science questions (GPQA Diamond) and advanced mathematics (AIME 2026, HMMT 2026) to legal reasoning (LEXam), instruction following (IFStruct), and document extraction (ExtractBench, MDPBench).

The findings also reframe the competitive narrative around AI leadership. While Chinese organizations collectively dominate submission volume and American labs receive much of the global media attention, a Korean Pre-AGI startup is quietly outperforming both on the metric that arguably matters most: actually winning.

Key Takeaways

• VIDRAFT holds first place on 10 of 48 Hugging Face official benchmarks — more than any other organization in the world as of October 5, 2026.
• QuantID's new real-time map consolidates all 48 Hugging Face official leaderboards into one live dashboard, revealing participation patterns, country distribution, and ranking gaps.
• Chinese organizations account for 55% of participating entities by submission count, but a Korean AI startup leads in benchmark victories.

Frequently Asked Questions

Q: How many Hugging Face official benchmarks does VIDRAFT currently lead?

A: As of October 5, 2026, VIDRAFT holds first place on 10 official Hugging Face benchmarks, including GPQA Diamond, MMLU-Pro, AIME 2026, and HMMT 2026, among others.

Q: What is QuantID's Hugging Face benchmark map, and how does it work?

A: It is a real-time interactive dashboard that reads Hugging Face's public API on every page load to display participation data, category breakdowns, country rankings, and first-place holders across all 48 official benchmarks — without storing or altering the source data.

Q: Which country has the most organizations participating in Hugging Face official benchmarks?

A: China leads with 55% of participating organizations, followed by the United States at 23% and South Korea at 4%, based on October 5, 2026 data from QuantID's map.

Source: 디지털타임스 (2026-10-06) — original article