First Snapshot of Hugging Face Official Benchmark Leaderboards: A Rundown of the 49 Leaderboard Champions

The First Snapshot of the Official Benchmark Map

The Hugging Face Official Benchmark Leaderboards map has released its first daily update. This data is read directly from the Hugging Face public API for all leaderboards tagged as "official benchmarks." As of 2026-10-18 (KST), it covers 49 official benchmarks and 1,102 leaderboard records.

Since this is the first installment in a series, its purpose is not "change tracking" but a "baseline record" — first answering one question: who is ranked first on each leaderboard today. Subsequent updates will report only what has changed. This means the snapshot itself has archival value: it is a time slice that can be referenced later to observe model iterations and leaderboard fluctuations.

Who Took the Most First Places

Looking at the dimension of "number of first-place finishes by organization," the leaderboards show a clearly uneven distribution.

Rank	Organization	First-place finishes
1	FINAL-Bench	11
2	inclusionAI	4
3	zai-org	4
4	XiaomiMiMo	3
5	deepseekAd-ai	3
6	MiniMaxAI	2
7	moonshotai	2
8	ornith-ai	2
9	Edge0	1
10	KDLAI	1

FINAL-Bench leads by a wide margin with 11 first places, but this number itself requires careful interpretation. The original article discloses at the end: QuantID is a technical partner of VIDRAFT, and FINAL-Bench is an organization under VIDRAFT that appears in these rankings. Therefore, this leading position should be viewed in light of the conflict-of-interest disclosure, rather than being directly equated with absolute technical superiority.

In the second tier are inclusionAI and zai-org, each with 4 first places. One detail worth noting: the first places of these two organizations are distributed across different types of tasks — inclusionAI's Ling-3.1-flash is concentrated on software engineering and skills benchmarks, while zai-org's GLM series covers scenarios such as research, evasion, and Chinese-language evaluation.

Top Models in Key Domains

On core capability benchmarks, several models stand out.

Knowledge and Reasoning

The top of MMLU-Pro (142 records, one of the leaderboards with the most records) is FINAL-Bench/Darwin-180B-RSI with a score of 88.12; it also tops gpqa with a score of 94.44. On the harder hle benchmark, deepseek-ai/DeepSeek-V4.1-Flash takes first place with 63.9.

Math benchmarks show perfect scores: on aime_2026 and hmmt_feb_2026, Darwin-180B-RSI scores 100 on both; the top of gsm8k is XiaomiMiMo/MiMo-V2.5-Pro with 99.6. Scores like these — near or at saturation — indicate that traditional math benchmarks are rapidly losing their ability to discriminate among current models.

Benchmark saturation: When multiple models approach or achieve perfect scores on a benchmark, that benchmark's ability to distinguish model capabilities drops sharply — a phenomenon known as benchmark saturation. Perfect scores on aime_2026 and hmmt_feb_2026 and 99.6 on gsm8k are typical signals of saturation. Saturation does not mean models have "solved" all problems in the field; rather, it means the difficulty distribution of the benchmark's questions can no longer cover the upper bound of current model capabilities. At this point, the community typically shifts to harder, newer, or more realistic benchmarks (such as hle, gpqa, etc.) to restore discrimination. Understanding saturation helps avoid misreading high scores: perfect scores reflect more the ceiling of the benchmark than the absolute capability boundary of the model.

Software Engineering and Terminal Tasks

The top of SWE-bench_Verified is ornith-ai/Ornith-1.5-397B with a score of 86; SWE-bench_Pro is taken by inclusionAI/Ling-3.1-flash with 68.74, and the same model also ranks first on SWE-bench_Multilingual (88.5) and skillsbench (68.7), demonstrating cross-language, cross-task stability.

Competition on terminal benchmarks is more fragmented: the top of terminal-bench-2.0 is mindlab-research/Macaron-V1-Venti (87.6), terminal-bench-2.1 belongs to DeepSeek-V4.1-Flash (90.6), while the base version terminal-bench is led by Ling-3.1-flash with 46.03.

Multimodal, Speech, and Document Parsing

In vision and multimodal, MMMU_Pro is topped by Darwin-180B-RSI with 79.48, and the top of ScreenSpot-Pro is meta-models/Muse-Glimmer-30B (75.4). In speech, the first place on open-asr-leaderboard is Edge0/ARK-ASR-3B with a score of 4.76 (lower is better on this leaderboard).

Document parsing is another highly competitive niche: ParseBench is led by KDLAI/KDL-Frontier-Parser-nano with 76.36, olmOCR-bench belongs to infly/Infinity-Parser2-Pro (87.6), and Real5-OmniDocBench is taken by PaddlePaddle/PaddleOCR-VL-1.6 with 93.19.

Agents and Long-Horizon Tasks

First places on agent benchmarks are fairly scattered: apex-agents is led by moonshotai/Kimi-K3 (41), Long-Horizon-Terminal-Bench is also Kimi-K3 (6), agents-last-exam belongs to XiaomiMiMo/MiMo-V2.6-Pro-RL (31.6), and Toolathlon is taken by the upgraded version of the same series, MiMo-V2.6-Pro-RL, with 76.9.

How to Correctly Interpret These Numbers

The original article offers two key reading notes, and any analysis citing this leaderboard should carry both.

First, the vast majority of leaderboard records are self-reported by model authors, and evaluation settings differ across submissions — sampling methods, voting strategies, thinking budgets, and more may all vary. Only entries explicitly marked by Hugging Face are "verified." This means cross-model score comparisons carry inherent uncertainty.

Second, when first places are tied, each tied organization is counted in the tally. Therefore, "number of first places" reflects the breadth of top finishes, not a strict capability ranking.

Combined with the conflict-of-interest disclosure mentioned earlier — the publisher QuantID has a partnership with the top organization FINAL-Bench, and the original article explicitly states that "this is not an independent evaluation" — this snapshot is better treated as an index of public data rather than an authoritative ranking.

The difference between self-reported and verified: On benchmark leaderboards, self-reported means the model author or organization proactively submits scores and configurations to the leaderboard, and the platform typically does not mandate verification; verified means the platform (such as Hugging Face) has performed some form of checks on the submission, for example confirming the evaluation script, data split, or runtime environment. When the two are mixed, readers can easily compare numbers of differing credibility on the same scale. Understanding this is essential to grasping why the original article emphasizes that "only marked entries count as verified" — unmarked scores should be treated as author claims, not independently reproduced results.

What inconsistent evaluation settings mean in practice: Sampling methods (greedy decoding vs. multi-sample voting), voting strategies (majority voting, weighted voting), thinking budgets (the tokens or compute a model may consume during the reasoning stage), and similar factors can all significantly affect final scores. For example, a model allowed a longer thinking budget may score higher on math or reasoning benchmarks, but this does not directly equate to stronger model capability. When comparing across models, if these configurations are not aligned, score differences may stem partly from evaluation protocols rather than model capability.

The Significance of This Baseline Map

Looking at the distribution of first places across these 49 leaderboards, several trends are worth noting.

"Small leaderboards" dense with proper nouns (fewer than 10 records) account for nearly half, indicating that the official benchmark ecosystem is rapidly fragmenting — new benchmarks keep appearing, but participation is still uneven. For example, arguana has only 3 records, and vlabench and BRIGHT have only 1 each.

That said, Chinese-background model organizations occupy the vast majority of top positions in this baseline — inclusionAI, zai-org, XiaomiMiMo, deepseek-ai, MiniMaxAI, and moonshotai together account for 18 first places. This distribution aligns with the activity of open-source models in the Hugging Face ecosystem in recent years.

The value of subsequent updates lies in change. When models iterate, new entries pour in, and top positions change hands, this baseline will become the reference point for judging "who is truly making progress." Until then, readers should treat every score as self-reported data with context, not as absolute proof of capability.

One-line Summary

Hugging Face's first official benchmark snapshot covers 49 benchmarks and 1,102 records, with Chinese model organizations taking 18 first places.

Paragraph Summary

The Hugging Face Official Benchmark Leaderboards map released its first daily update, covering 49 official benchmarks and 1,102 leaderboard records as of 2026-10-18, positioned as a baseline for tracking future changes. FINAL-Bench leads with 11 first places, but the publisher QuantID has a partnership with it, and the original article explicitly states it is not an independent evaluation. Chinese-background model organizations together take 18 first places, with inclusionAI's Ling-3.1-flash performing strongly on multiple software engineering benchmarks. aime_2026 and hmmt_feb_2026 show perfect scores of 100, and gsm8k reaches 99.6, indicating that traditional math benchmarks are saturating in discriminative power. Most leaderboards are self-reported by authors with inconsistent evaluation settings; only entries marked by Hugging Face count as verified, so cross-model score comparisons carry inherent uncertainty.