Qwen Still Takes Half of Open-LLM Downloads: What Changed in the Hugging Face Top 300 Since September

TL;DR

Is Qwen losing its grip on open-model downloads? A little, and less than the headline number suggests. We compared two snapshots of the 300 most-downloaded text-generation models on Hugging Face, taken on September 17 and November 4, 2026, ranked by trailing 30-day downloads. Qwen's share of the Top 300 fell from 55.8% to 50.9%. Still, one family out of dozens accounts for half of all downloads in the table.

Three findings matter more than the headline:

• One 2022 model distorts the totals. openai-community/gpt2 went from 3.8% to 8.7% of Top-300 downloads. With it removed, Qwen's share only slips from 58.0% to 55.8%.
• The table turns over fast. 106 of the 300 rows were not in the September list, and 71 of those were created in 2026. Even so, the top 10 hold more than before (28.0% to 32.5%).
• Quantized formats keep growing. FP8 went from 6.6% to 8.3% and GGUF from 15.8% to 17.4%, while NVFP4 and MXFP4 both slipped.

Disclosure: VIDRAFT builds and maintains the leaderboard these snapshots come from, and some VIDRAFT models appear in the Korean section. The method is public and the code below reproduces the core of it.

What exactly did we measure?

The Global LLM Download Leaderboard ranks Hugging Face text-generation repositories by trailing 30-day downloads, not cumulative downloads. Cumulative totals reward age. A 30-day window shows what people are pulling right now. Each row also carries cumulative downloads, so you can see both.

Two rules shape the numbers:

1. Family is attributed by architecture, not by name. A third-party quantization counts toward the family that trained the original, resolved from the model's model_type. Names can be changed and architectures cannot.
2. Non-models are removed. CI fixtures such as tiny-random-* and embedding or reranker repositories are excluded, because they pile up downloads without being models anyone chats with.

Here are the two snapshots side by side.

The total fell by 13.6%. We cannot say why from this data alone, so every comparison below uses shares of each snapshot's own total rather than raw counts, except where we say otherwise.

Is Qwen losing the open-model market?

Not in any sense a Qwen team would worry about. Qwen's share dropped 4.9 points, from 55.8% to 50.9%. Its row count barely moved (139 rows in September, 136 in November). The Qwen organization alone, counting only its own repositories, accounts for 34.5% of all Top-300 downloads on November 4. Four of the five most-downloaded rows after gpt2 are Qwen's own Qwen3 checkpoints.

The absolute numbers tell a slightly different story. Qwen-family downloads in the table fell from 220.3M to 173.7M, a drop of 46.6M, or 21%. That is a larger fall than the 13.6% drop in the table overall, so the share loss is real and not just a side effect of a shrinking pool.

The losers and winners are not symmetric. Gemma lost the most share, from 5.6% to 2.5% (22.3M down to 8.7M downloads). Llama slipped from 6.5% to 5.3%. On the other side, DeepSeek rose from 4.6% to 6.7%, GLM from 2.1% to 3.4% and Nemotron from 2.6% to 3.2%. A further 8.1M downloads moved into rows the leaderboard could not attribute to any family, which is a reminder that attribution is a measurement with an error bar, not a fact of nature.

By country of the training family, Chinese-origin models went from 66.1% to 64.1% of downloads and US-origin from 31.5% to 30.0%. The difference was absorbed mostly by the unattributed bucket (0.7% to 3.9%) and EU families (1.2% to 1.6%). Korean-origin models held at 0.5% in both snapshots.

Why does one 2022 model skew the totals?

Look at rank 1 in the table: openai-community/gpt2, with 29.8M downloads in 30 days and about 963M cumulative. It is a 0.14B-parameter model released in 2022. It already had 15.2M in September. Its downloads nearly doubled, and its share of the Top 300 more than doubled, from 3.8% to 8.7%.

We do not believe 30 million people deployed gpt2 last month. Our best reading is automated traffic: tutorials, test suites, CI jobs and libraries that fetch a small, always-available model. We cannot separate that traffic from human use, and Hugging Face's own counter does not either. That is why the leaderboard documents download counts as a measure of distribution, not users.

The effect on the family table is large because gpt2 sits in the GPT/OPT family. Remove it and the picture changes:

With gpt2 in, GPT/OPT looks like it gained 3.3 points. With gpt2 out, it lost 1.3. If you read the first column of any "open model share" chart without checking for a single outlier, you can end up with the opposite conclusion. Qwen's decline also shrinks by more than half, from 4.9 points to 2.2.

Are people moving to smaller, quantized models?

Partly. By parameter count, models under 3B grew from 27.9% to 30.9% of downloads, and models of 70B or more grew from 9.9% to 12.2%. The middle shrank: 3B to 10B fell from 28.4% to 24.6%, and 10B to 70B from 30.1% to 27.3%. Both ends gained, which fits a split between laptop and edge use on one side and very large mixture-of-experts checkpoints on the other.

The 70B+ group changed composition as well. In September its top rows were openai/gpt-oss-120b (5.4M), deepseek-ai/DeepSeek-V4-Flash-0731 (4.5M) and Qwen/Qwen-72B (3.7M). In November the leader is deepseek-ai/DeepSeek-V3.2 at 6.9M, a 685B-parameter FP8 checkpoint that had 2.0M in September.

For quantization format, the picture is this:

Full-precision rows still make up about 60% of downloads in both snapshots. Among compressed formats, GGUF and FP8 grew, and NVFP4 (6.1% to 4.4%) and MXFP4 (3.2% to 2.5%) shrank. The FP8 rows are dominated by large DeepSeek and GLM checkpoints rather than small models. The leaderboard's type tag moved too: "Instruct" fell from 24.5% to 14.3% and "Base" rose from 35.6% to 45.1%. We would not lean on that last one, because the type label comes from a metadata and name classifier and gpt2 is a "Base" row.

How fast does the Top 300 turn over?

Faster than most people assume. Compared with the September list, 106 rows are new on November 4 and 106 dropped out. The new rows hold 18.5% of November downloads. Seventy-one of the 106 were created in 2026, and the median creation date among them is July 11, 2026.

The biggest newcomer is ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF, a GGUF repack created on October 23 that already pulled 2.5M downloads in 30 days. It is a good example of why the leaderboard attributes by architecture: by name it looks like a one-off community upload, by architecture it belongs to the DeepSeek family and counts there.

Concentration moved the other way. The top 10 rows went from 28.0% to 32.5% of downloads, and rank 1 grew from 20.2M to 29.8M while rank 300 shrank from 272K to 219K. The long tail is churning, and the head is getting heavier.

Where does Korea sit in this picture?

Korean-origin families hold 0.5% of Top-300 downloads in both snapshots. The Korea Top 30 section lists Korean organizations separately. Its leader on November 4 is FINAL-Bench/POCKET-35B-GGUF (a VIDRAFT model) at 531K downloads, down from 815K in September, with POCKET-Darwin-180B-GGUF entering at 398K. Next come LG AI Research's EXAONE 3.5 7.8B Instruct (366K) and its AWQ build (183K). Because VIDRAFT models lead that section, read this paragraph as a disclosure first and a finding second.

How can I reproduce the core of this in 20 lines?

The snippet below uses the public Hugging Face API through huggingface_hub. It groups the current Top 300 by model_type, the raw architecture field our family attribution starts from. It will not match our table exactly, because our attribution also resolves GGUF and other repackaged repositories back to the base model they came from.

from collections import Counter
from huggingface_hub import HfApi

api = HfApi()
rows = list(api.list_models(
    pipeline_tag="text-generation",
    sort="downloads",          # trailing 30-day downloads
    limit=300,
    expand=["downloads", "createdAt", "config"],
))

share = Counter()
for m in rows:
    model_type = (m.config or {}).get("model_type", "unknown")
    share[model_type] += m.downloads or 0

total = sum(share.values())
for model_type, d in share.most_common(8):
    print(f"{model_type:12s} {d / total:6.1%}")
print(len(rows), total)

We ran it on November 5, 2026. The total came to 339.0M downloads across 300 rows, within 1% of our November 4 snapshot (341.1M). The top of the output:

qwen3         22.3%
unknown       16.5%
qwen2         12.1%
gpt2           9.9%
llama          8.0%
nemotron_h     3.1%
qwen3_5_moe    2.8%
gpt_oss        2.7%

Two things are worth reading off this. First, unknown is 16.5% because GGUF repositories usually carry no config.json, so a raw model_type count cannot place them. That is exactly the gap family attribution is built to close. Second, qwen3, qwen2 and qwen3_5_moe are three separate keys here but one family in our table, so the raw output understates Qwen. Treat it as a quick check on direction, not a replacement for the table.

What do download counts not tell you?

They do not count users, API calls, private deployments or models distributed through other channels. A report in The Next Web on August 16, 2026 made the same point from the other side: Alibaba's own figure of roughly 3 billion Qwen downloads is about 1.5 times the 2.045 billion that Hugging Face counted for the year. Hugging Face notes that its count covers activity inside its own hub only, and Alibaba also distributes through its own ModelScope platform. The two numbers measure different things.

For our snapshots, three limits apply:

• Hugging Face only. Nothing here covers other model hubs.
• Automated traffic is included. The gpt2 case is the clearest, but it is probably not the only one.
• Attribution is imperfect. Roughly 3% of downloads sit in rows we could not attribute to a family, and the country view carries 3.9% as Other.

Two snapshots 48 days apart also cannot separate a trend from a bump. We keep refreshing the table daily and would rather publish a third point than argue about two.

FAQ

Which model family has the most open-model downloads on Hugging Face in 2026?
Qwen. In our November 4 snapshot of the Top 300 text-generation models, Qwen accounts for 50.9% of trailing 30-day downloads, down from 55.8% on September 17.

Why is gpt2 the most downloaded text-generation model?
We cannot prove the cause, but the likeliest explanation is automated traffic from tutorials, test suites and libraries that load a tiny model by default. It reached about 963M cumulative downloads, and it held 8.7% of Top-300 downloads on November 4.

Do Hugging Face download counts measure real usage?
No. They measure distribution on one hub, including automated and mirrored traffic. They do not include API usage, private deployments or other model hubs.

Is quantization becoming more common among popular open models?
Compressed formats make up about 40% of Top-300 downloads in both snapshots. Within that, FP8 (6.6% to 8.3%) and GGUF (15.8% to 17.4%) grew, while NVFP4 and MXFP4 shrank.

How is a model's family or country decided?
By the architecture (model_type) of the model that was trained, so a third-party quantization counts toward the original trainer. Rows without enough evidence stay "Other" and are not guessed.

Further reading

• VIDRAFT POCKET: Running a 180B MoE Model on an 8 GB VRAM Laptop
• POCKET-35B-GGUF: Running a 35-Billion-Parameter LLM on CPU Without a GPU

Data: Hugging Face API, snapshots of September 17 and November 4, 2026, plus the November 5 check above. Leaderboard and methodology (CC BY 4.0): VIDraft/global-llm-leaderboard. Model names are trademarks of their respective owners.