We ranked the top 300 open LLMs by real downloads. 66% of them are Chinese.
Every "top open model" list I could find ranks by cumulative downloads. Cumulative rewards age: a
model from 2022 outranks one that is actually being used today. So we built the other view —
trailing 30-day downloads on Hugging Face text-generation, global top 300 plus Korea top 30 —
and published it.
The ranking is the boring part. The interesting part was deciding which country a model belongs
to, and that turned out to be where most of the work went.
The attribution problem
Take unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF. Twelve million downloads in 30 days. Whose model
is that?
The uploader is unsloth. The model is Qwen. If you attribute by uploading organisation, you get a
picture of who repackages, not of who trains. Quantized re-uploads are a large share of all
download volume, so this choice moves the result substantially.
We attribute by model family — whoever trained the original. A third-party quantization of Qwen
counts as China.
Attribute by architecture, not by name
Our first pass resolved families from repository names. That failed in a specific way. Names get
changed; a post-trained derivative can be called anything. We had 19.7% of the top 300 unattributed
— worse than the previous run, because the newly-charting repositories were exactly the ones
whose names told us nothing.
Architecture is harder to disguise. model_type in the config says what the thing actually is:
ornith-ai/Ornith-1.0-35B → qwen3_5_moe
ornith-ai/Ornith-1.5-397B → qwen3_5_moe
prism-ml/Bonsai-27B-mlx-1bit → qwen3_5
Resolving by architecture, then letting quantized variants inherit from their non-quantized sibling,
took unattributed from 19.7% → 2.7%.
Two things we got wrong on the way, both worth naming:
-
A tool name is not a family name. Our matcher hit
llamainsidellama.cppin a GGUF tag and attributed three repositories to Llama.llama.cppis a converter, not a lineage. -
CI fixtures were still in the top 300.
tiny-gpt2,tiny-random-Llama-3, and friends accumulate real download counts from automated test suites. Four of them were ranking. They are not models anyone uses.
Whatever we could not resolve from architecture stayed Other. We did not guess. The unresolved
share is printed in the page footnote, because a leaderboard that hides its own coverage gap is
asking to be believed on faith.
What the data says
Top 300 by trailing 30-day downloads, share of total volume:
| Share | |
|---|---|
| China | 66.1% |
| United States | 31.5% |
| EU | 1.2% |
| Korea | 0.5% |
| Other | 0.7% |
Narrow to GGUF — the format people use to run models locally, without a GPU — and 65 of the top
300 qualify: 47 Chinese, 12 American, 2 Korean.
That GGUF slice is the one I would watch. It is the closest proxy in this data for someone ran this
on their own machine, and it is even more concentrated than the overall picture.
Disclosure: two of those rows are ours
We build open models, so we are in our own leaderboard. Stating it plainly:
-
FINAL-Bench/POCKET-35B-GGUF— #103 overall, #13 among GGUF models -
FINAL-Bench/POCKET-26B-GGUF— #231
POCKET-35B passed 1,076,570 cumulative downloads about seven weeks after release, and 76% of
that cumulative total landed in the last 30 days — which is why it ranks where it does on a 30-day
axis and much lower on a cumulative one. Both numbers are on the page.
Those two rows are 60.2% of all Korean download volume in the top 300. Korea's total is 0.5%. Both
of those facts are true at the same time, and the second one is the more important one.
Limits, stated up front
Download counts are not users. They are what the platform reports, and they include automated
pulls and mirror traffic that we cannot separate. This measures distribution. Anyone quoting it
as adoption is over-reading it, us included.
It is a snapshot, not a feed. The measurement date is printed at the top of the page and in
every export. We refresh it manually.
2.7% is unattributed. Self-designed architectures with no public lineage — we left them as
Other rather than inventing an answer.
The page is open and the methodology is CC BY 4.0:
huggingface.co/spaces/VIDraft/global-llm-leaderboard
If you find a row attributed wrongly, the architecture field is the thing to check first — that is
where our resolution comes from, and it is the part most likely to be wrong in an interesting way.
Top comments (0)