Global LLM Download Analysis: Chinese Models Dominate 70.7% of Top 300 on Hugging Face — One Korean Model Makes the Cut

TL;DR: VIDRAFT analyzed the top 300 most-downloaded language models on Hugging Face over a 30-day window, covering 523 million total downloads across 6 dimensions: country of origin, model family, release date, format, parameter size, and publishing account. The data reveals extreme concentration — a single Chinese model family accounts for 60.3% of all downloads — and only one Korean model, VIDRAFT's own POCKET-Darwin-180B, cracked the top 300. Developers tracking global LLM adoption trends or benchmarking their own open-source releases will find the methodology and leaderboard directly actionable.

What it is

VIDRAFT (비드래프트), a Korean AI deep-tech startup, published a structured analysis of Hugging Face's public download API data, captured on November 17, 2026. The study examines the top 300 language models by 30-day download count — totaling 523.22 million downloads — sliced across six analytical axes:

  • Country of origin (where the base model was first developed)
  • Model family (e.g., Qwen, DeepSeek, Gemma, LLaMA)
  • Release date (year the model was first published)
  • Format (original weights vs. quantized variants like GGUF, FP8, NVFP4)
  • Parameter count (model size buckets)
  • Publishing account (original developer vs. third-party re-uploader)

The resulting leaderboard is published on Hugging Face and updates daily.

How it works

The methodology is transparent and reproducible by any developer with Hugging Face API access:

  1. Data collection: Query Hugging Face's public API for the language model task category, sorted by 30-day download count. Snapshot taken on 2026-11-17.
  2. Filtering: Vision-only models (image captioning, OCR-exclusive) were excluded. Only language models were counted.
  3. Attribution: Country and family labels are assigned based on the original developing organization, not the account that uploaded the specific file. This matters because 29.6% of downloads come from third-party re-uploads (quantized/derived forks) rather than the original developer's account.
  4. Comparison cohort for Korean companies: Restricted to models published on official company accounts after January 1, 2026, covering 11 representative Korean AI organizations.

The leaderboard refreshes daily from the same public API, so the rankings are a live signal of real developer pull — not curated lists or survey responses.

Benchmarks & results

Global picture (top 300, 30-day window as of 2026-11-17):

  • China: 206 models, 70.7% share
  • USA: 76 models, 26.8% share
  • Europe: 4 models, 0.6% share
  • Korea: 1 model, 0.1% share
  • Total downloads across all 300: 523.22 million

By model family:

  • Alibaba's Qwen family (including derivatives and quantized versions): 173 of 300 models, 60.3% of all downloads
  • GPT/OPT family (GPT-2, gpt-oss, OPT, etc.) and Google Gemma family: 10.5% each
  • DeepSeek family: 4.9%
  • Zhipu GLM family: 2.9%
  • Meta LLaMA family: 2.2%
  • NVIDIA Nemotron family: 1.8%

Top individual models (30-day downloads):

  1. GPT-2 (released 2019) — 34.19 million (widely used for education and software testing)
  2. Qwen3 0.6B — 17.07 million
  3. Qwen3 14B — 13.08 million
  4. DeepSeek V3.2 — 9.20 million
  5. 8 of the top 10 are Chinese-origin models
  6. Google Gemma 4 31B reached 6.35 million, placing 10th alongside GPT-2 as the only US-origin models in the top 10

2026 models specifically:

  • 197 of 300 models were released in 2026, accounting for 56.7% of downloads
  • 2025 models: 22.9%; 2024 models: 11.0% — recency strongly predicts rank

Formats:

  • Quantized formats (GGUF, FP8, NVFP4): 49.1% of all downloads
  • GGUF alone (runs on laptops/PCs): 27.5%

Model sizes: 15B–70B "mid-size" models led at 32.7%; 3B–15B at 29.0%; under 3B at 21.1%; 70B+ large models accounted for 9.4% (33 models in the top 300).

Korean company comparison (2026 models only, 11 companies):

Company Models published 30-day downloads Best global rank
VIDRAFT 99 1.12M (82.2% of Korean total) #260 (POCKET-Darwin-180B)
SK Telecom 5 86K #867
Krafton 9 66K #1,208
LG AI Research 11 50K #1,852
Kakao 9 23K —

VIDRAFT's POCKET-Darwin-180B — a heavily quantized version of Darwin-180B (180 billion parameters) designed to run on a single 128 GB RAM PC or laptop without datacenter GPUs — is the only Korean model in the global top 300 at #260 with 528,000 downloads in 30 days. The base model Darwin-180B ranked #1 in math and science reasoning among open-source models in SuperCLUE's November 2026 evaluation.

How to try it

The Global LLM Download Leaderboard that VIDRAFT publishes from this analysis is publicly available on Hugging Face and updates daily — search for VIDRAFT's leaderboard on huggingface.co.

POCKET-Darwin-180B is publicly available on Hugging Face. You can browse and download it via the Hugging Face Hub:

huggingface-cli download VIDRAFT/POCKET-Darwin-180B 

Note: Confirm the exact repository slug on the VIDRAFT Hugging Face profile before running. The 180B quantized model is sized for 128 GB RAM environments.

FAQ

Q: Why does GPT-2 (a 2019 model) still rank #1 by downloads despite being years old?
A: The analysis attributes this to its near-universal use as a baseline in education, ML coursework, and software testing pipelines. Its small size, permissive license, and ubiquity in tutorials mean it gets pulled programmatically at massive scale — a good reminder that "most downloaded" ≠ "most capable."

Q: Nearly 30% of top-300 downloads come from third-party re-uploads — what does that mean for attribution?
A: Accounts specializing in quantization (like compression-focused community uploaders) redistribute original models in GGUF or FP8 formats optimized for consumer hardware. VIDRAFT's methodology attributes the model family and country to the original developer, but counts downloads regardless of which account served the file. This means adoption numbers for popular families like Qwen may actually be understated if some derivative uploads weren't captured.

Q: How can I replicate this analysis myself?
A: The underlying data comes from Hugging Face's public API — query the language model task category sorted by 30-day downloads, then apply the six classification axes VIDRAFT used. The daily-refreshed leaderboard on VIDRAFT's Hugging Face page also gives you a live baseline to compare against.

Originally reported by 전자신문 (2026-11-18) — source article.