The Top Spot in AI Model Downloads Isn’t a Chatbot

Picture the open AI models everyone talks about. DeepSeek’s reasoning model, Meta’s Llama, Alibaba’s Qwen, the image generator FLUX. Those names come up first. Now count the models people actually pull, and the list looks nothing like that.

On July 21, 2026, I tallied the top 100 models by downloads on the Hugging Face Hub directly. The number-one model wasn’t a chatbot. It was all-MiniLM-L6-v2, a 22.7-million-parameter embedding model that logged more than 240 million downloads in the trailing 30 days. The ten most-hyped open LLMs, added together, don’t reach one-sixth of that.

The AI model downloads chart belongs to the retrieval and representation plumbing running quietly underneath the headlines.

How I measured it

The measurement ran against Hugging Face’s public REST API. From the huggingface.co/api/models endpoint, which needs no authentication, I collected the top 100 models by downloads and the top 40 by likes, then queried ten of the most-hyped generative LLMs individually. Each model’s category came not from my judgment but from the pipeline_tag Hugging Face assigns, mapped straight through.

One field definition matters here. The downloads figure is the trailing 30 days, not all-time. Likes measure human attention, and downloads measure machine usage. They are two different signals. To size the gap between them, I also computed downloads-per-like (downloads/likes). The commands, raw JSON, and aggregates are preserved so the run reproduces.

No chatbot in the top 9 of AI model downloads

Here is the download top 10 as I tallied it. Ranks one through nine are all non-generative models. A generative chat model appears only at #10, and even that is not a large conversational model but Qwen3-0.6B, a 0.6-billion-parameter part that typically gets baked into test and automation pipelines.

Grouping by category sharpens the picture. Of the top-100 downloads, 35.9% were embeddings, 8.7% rerankers, and 12.2% BERT-style classic NLP. That retrieval-and-representation infrastructure sums to 56.8% of the volume, while generative chat LLMs came to just 14.3%. Chat models were the most numerous category by headcount, with 23 of the top 100, yet their individual downloads were small enough that, in aggregate volume, they trailed the embedding category alone.

The result lines up with a community analysis published on the Hugging Face blog. Across the Hub, the 50 most-downloaded entities take 80.22% of all downloads, and 92.48% of downloads go to models under one billion parameters. Small, repeatedly pulled models, not the giant generators, produce most of the download volume.

The likes ranking points the other way

Now sort the same data by likes. The likes leader is the image generator FLUX.1-dev with 13,720, but it logged only about 573,000 downloads in 30 days. DeepSeek-R1 sits right behind at 13,490 likes, second on the Hub’s overall likes chart, yet its downloads still run at roughly one-twenty-eighth of the embedding leader.

Lay the two lists side by side and almost no names repeat. Likes go to the generative models people find impressive enough to star. Downloads go to the retrieval and representation models pipelines actually pull. The one standout exception is Kokoro-82M, a tiny speech model at likes rank six that also logged 9.7 million downloads, and it proves the same rule from another angle, since small models that get embedded in pipelines rack up pulls. Set that exception aside and the two rankings point in different directions because they are different signals.

Measuring the gap with downloads-per-like

To compress that gap into one number, I computed downloads-per-like. all-MiniLM, the archetype of pure infrastructure, scores 47,185. FLUX.1-dev, the emblem of hype, scores 41.7. On the same metric, that’s roughly a 1,130x spread.

Why embeddings dominate downloads

There’s a structural reason embeddings and rerankers own the download charts. They are core parts of retrieval-augmented generation (RAG) and vector search pipelines. Turning documents into vectors, storing them, and fetching the closest chunk to a query all route through an embedding model, and a reranker often follows to sharpen the results.

all-MiniLM converts a sentence into a 384-dimension vector in a roughly 90MB package that is small, fast, and tuned for semantic search. Search, chatbot, and recommendation services each deploy it, and every deployment has CI and automation pull it again. Large generative LLMs, by contrast, are usually called through a cloud API or downloaded once and kept, so their download counters climb more slowly.

External data points the same way. In Analytics Vidhya’s 2025 list of the most-downloaded open-source models, MiniLM again ranked first, followed by small encoders like ELECTRA and DistilBERT, which means this isn’t a quirk of one snapshot but a structurally recurring pattern.

What downloads reveal about the real AI stack

The tally points to where production AI’s center of gravity actually sits. While the narrative fixates on the performance race between generative models, a large share of the stack shipping every day is the infrastructure handling search and representation. Like the gap between agent benchmark scores and real-world performance, mistaking hype for usage means missing where the actual deployment weight lands.

Will this picture hold next quarter? I’ll keep watching whether retrieval infrastructure (embeddings, rerankers, and classic NLP) keeps holding half of the top-100 downloads, and whether any generative chat model’s individual downloads start closing on the embedding leaders. Those two lines, not the buzz, will show where production AI’s real weight is moving.

References

• Hugging Face – Models sorted by downloads (2026-07-21) – data source for the direct top-100 downloads and top-40 likes tally
• Hugging Face Blog (lbourdois) – Model statistics of the 50 most downloaded entities (2026) – community write-up, top 50 = 80.22% of all downloads, 92.48% under one billion parameters
• Sentence Transformers Documentation – Pretrained Models (2026) – official model overview, all-MiniLM-L6-v2 recommended as about five times faster than the top-quality all-mpnet
• Analytics Vidhya – Top Open-Source Models on HuggingFace (2025-11) – external corroboration that MiniLM led 2025 downloads, with ELECTRA and DistilBERT near the top
• Hugging Face – all-MiniLM-L6-v2 model card (2026-07-21) – model card of the number-one model (384-dimension embeddings, 22.7M parameters, cumulative downloads)

jugger

Jugger | Backend Developer

A backend engineer working in enterprise solutions software since June 2021. The hands-on perspective from system design, API implementation, and production operations directly informs this site’s
AI, IT, and cloud infrastructure coverage.

For topics outside direct professional experience—blockchain and macroeconomics—analysis is built on primary-source citations such as SEC filings, whitepapers, institutional research, and government
statistics. Speculation is clearly separated from verified fact, and unverified claims are not published.

AI tools assist with drafting, but every article is reviewed, revised, and verified before publication. Errors are corrected with visible notes when discovered.

Contact: jjunggulee@gmail.com