Hugging Face Official Leaderboard Daily Update: 49 Benchmarks, SoMarkAI Tops Visual Parsing

Hugging Face's official benchmark leaderboard ecosystem is expanding rapidly. According to the latest snapshot of the leaderboard map maintained by QuantID, as of 2026-10-20 (KST), the platform has aggregated 49 official benchmarks with 1,107 leaderboard results recorded in total. This number itself is more noteworthy than the ranking of any single model: it shows that the community's demand for "reproducible, comparable" evaluation has spread from a handful of star leaderboards to a matrix structure covering OCR, Agent, code, legal, speech, and other specialized domains.

First, a note on how to read these leaderboards: the vast majority of results are self-submitted by model authors, with varying evaluation settings such as sampling parameters, voting strategies, and thinking budgets. Only entries marked by Hugging Face as official are verified. Therefore, the rankings below are better viewed as "an aggregated view of community self-reported results" rather than independent third-party evaluation.

Biggest Change This Period: olmOCR-bench Has a New King

The core changes over the past day centered on the document parsing track. SoMarkAI's OXR-1.0 climbed to the top of olmOCR-bench with a score of 87.84, replacing the previous infly/Infinity-Parser2-Pro. The leaderboard currently has 22 result records, making it one of the most competitive battlegrounds in the OCR space.

During the same period, 5 new leaderboard results were added across different benchmarks—a modest increment—but a change at the top often signals the start of a new wave of submissions. Document parsing has been one of the areas with the strongest deployment demand over the past year—structured extraction from invoices, contracts, scanned documents, and multilingual tables all depends on this capability, giving leaderboard rankings more direct commercial significance than purely academic benchmarks.

olmOCR-bench is a document parsing evaluation benchmark introduced by the Allen Institute for AI (AI2), derived from the evaluation framework accompanying its open-source OCR model olmOCR. It is designed for real-world PDF scenarios, covering challenging units such as tables, formulas, multi-column layouts, headers and footers, and handwriting. It requires models to output structured text (such as Markdown) and then scores using normalized edit distance, table TEDS, and other metrics. Unlike traditional OCR, which only looks at character recognition rates, olmOCR-bench places greater emphasis on layout reconstruction and reading order, making it closer to the real needs of downstream tasks such as RAG and document QA. This is also why it has become the core arena of the document parsing track.

Organization Leaderboard: FINAL-Bench Leads by a Wide Margin with 11 First-Place Finishes

When counted by "how many leaderboard first places an organization holds," the gap between first place and the rest is striking:

Rank	Organization	Number of First Places
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

FINAL-Bench's 11 first-place finishes come from its Darwin-180B-RSI series models, covering MMLU-Pro (88.12), gpqa (94.44), MMMU_Pro (79.48), the LEXam series, hmmt_feb_2026 (100 points), and other areas. It must be noted, as disclosed in the original text: FINAL-Bench has a technical partnership with QuantID, the maintainer of the leaderboard, and its parent company VIDRAFT appears in the rankings. Therefore, this daily report is not independent evaluation, and readers should treat the related rankings with some caution.

Close behind, inclusionAI and zai-org each hold 4 first places. The former scores across multiple fronts with Ling-3.1-flash on SWE-bench_Pro (68.74), SWE-bench_Multilingual (88.5), and skillsbench (68.7), while the latter's GLM series occupies positions on specialized evaluations such as ResearchClawBench, EvasionBench, and chi-bench.

Top Models Across Tracks

Breaking down the 49 leaderboards by capability domain makes the top-level landscape clearer.

General Knowledge and Reasoning

MMLU-Pro and gpqa are both led by FINAL-Bench/Darwin-180B-RSI, while aime_2026 and hmmt_feb_2026 achieve perfect scores of 100 (these two math benchmarks have 27 and 20 records respectively—small sample pools, so perfect scores are of limited persuasiveness). On the hle benchmark, deepseek-ai/DeepSeek-V4.1-Flash leads with 63.9, a score that itself shows that high-difficulty reasoning sets remain tough nuts to crack.

Code and Terminal Agents

SWE-bench_Verified is taken by ornith-ai/Ornith-1.5-397B with 86, while SWE-bench_Pro goes to inclusionAI/Ling-3.1-flash (68.74). Terminal benchmarks diverge: terminal-bench-2.0 is led by mindlab-research/Macaron-V1-Venti (87.6), while terminal-bench-2.1 and deep-swe are both led by DeepSeek-V4.1-Flash. Different versions of leaderboards in the same capability domain producing different winners precisely exposes how evaluation version differences affect conclusions.

Multimodal, Speech, and Vertical Domains

ScreenSpot-Pro is led by meta-models/Muse-Glimmer-30B with 75.4 on GUI grounding tasks; on Real5-OmniDocBench, PaddlePaddle/PaddleOCR-VL-1.6 achieves 93.19; in speech, Edge0/ARK-ASR-3B tops open-asr-leaderboard with a WER of 4.76. The legal vertical benchmark LEXam-hard's top score is only 45.72—one of the lowest absolute scores in the entire table—showing that specialized domains remain a weakness for models.

A few other readings deserve attention: the top score on yc-bench is 2,100,000, and on ResearchClawBench it is 20.709230769230768. Raw, unnormalized scores like these are not recommended for cross-leaderboard comparison.

The SWE-bench series deserves a separate note on version differences. SWE-bench_Verified is a subset of 500 high-quality GitHub issues manually selected from the original SWE-bench, excluding samples with unclear descriptions or unreliable tests, making it the most widely cited benchmark in the code-fixing direction; SWE-bench_Pro targets more complex multi-file, long-chain engineering tasks and is significantly harder. Similarly, terminal-bench versions 2.0 and 2.1 differ in task sets, scoring scripts, or environment configurations, so rankings cannot be directly interchanged—which explains the phenomenon of "same capability, different winners."

WER (Word Error Rate) is the most commonly used metric in speech recognition, calculated as (substitutions + deletions + insertions) divided by the total number of words in the reference text—lower is better. It directly reflects the deviation between transcription output and human-annotated text, but is sensitive to homophones, punctuation, and number formats, so differences in text normalization across leaderboards affect comparability. open-asr-leaderboard uses WER as its primary metric; a score of 4.76 means about 4.76 words out of every 100 contain errors—already a very strong open-source level.

What's Really Worth Reading in This Daily Report

Individual rankings fluctuate with submissions, but structural signals are more stable. First, the number of leaderboards has reached 49, and long-tail leaderboards (those with fewer than 10 results) account for a significant share, meaning that many specialized domains have just entered the stage of "someone is evaluating but samples are insufficient"—early submissions have high marginal returns. Second, Chinese teams hold the majority of seats in the top ten organizations: DeepSeek, Zhipu (zai-org), MiniMax, Moonshot AI, Xiaomi MiMo, and Ant Group's inclusionAI all have multiple first-place finishes, showing a clear eastward shift in the competitive center of gravity for open-source models. Third, the self-reported mechanism means leaderboards are "signals" rather than "verdicts"—cross-leaderboard, cross-version validation is more valuable than fixating on a single first place.

One-line Summary

Hugging Face's official leaderboard has aggregated 49 benchmarks and 1,107 results; the OCR track has a new leader, and FINAL-Bench leads with 11 first places despite a disclosed relationship.

Paragraph Summary

As of 2026-10-20, Hugging Face's official benchmark leaderboard has aggregated 49 official benchmarks with 1,107 cumulative results, including 5 new entries that day. The biggest change is SoMarkAI's OXR-1.0 topping olmOCR-bench with 87.84, replacing infly/Infinity-Parser2-Pro. On the organization leaderboard, FINAL-Bench leads by a wide margin with 11 first places, but its technical partnership with leaderboard maintainer QuantID warrants caution; inclusionAI and zai-org each hold 4 first places, and Chinese teams make up the majority of the top ten. Top positions are dispersed across tracks: SWE-bench_Verified goes to Ornith-1.5-397B, different terminal-bench versions have different winners, and the legal benchmark LEXam-hard's top score is only 45.72, showing specialized domains remain a weakness. The article emphasizes that most leaderboard results are self-reported by authors with inconsistent evaluation settings, and cross-validation is more valuable than a single first place.