What the World Is Actually Downloading from Hugging Face in May 2026
Hugging Face is the dominant model registry for open-weight AI. The all-time download counters on its public model API are the most direct signal of which open-weight models developers are actually pulling into production, fine-tuning pipelines, and research workflows. This page ranks the most-downloaded models on Hugging Face as of May 14, 2026, broken into "everything" and "text-generation LLMs only." The result is sharply different from the consumer narrative around frontier closed models.
Top 15 Models on Hugging Face by All-Time Downloads (Any Task)
| Rank | Model | Task | Downloads | Likes |
|---|---|---|---|---|
| 1 | sentence-transformers/all-MiniLM-L6-v2 | Embeddings (sentence-similarity) | 259,230,702 | 4,786 |
| 2 | Qwen/Qwen3-VL-2B-Instruct | Multimodal (image-text-to-text) | 183,999,222 | 403 |
| 3 | google-bert/bert-base-uncased | Classic NLP (fill-mask) | 63,940,780 | 2,651 |
| 4 | google/electra-base-discriminator | Classic NLP | 52,284,659 | 105 |
| 5 | cross-encoder/ms-marco-MiniLM-L6-v2 | Re-ranking (text-ranking) | 50,880,509 | 238 |
| 6 | sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | Multilingual embeddings | 47,885,224 | 1,226 |
| 7 | BAAI/bge-small-en-v1.5 | Embeddings | 42,635,548 | 458 |
| 8 | sentence-transformers/all-mpnet-base-v2 | Embeddings | 35,940,012 | 1,290 |
| 9 | openai/clip-vit-large-patch14 | Vision (zero-shot) | 32,486,672 | 2,010 |
| 10 | BAAI/bge-m3 | Embeddings (multilingual) | 24,970,361 | 2,997 |
| 11 | openai/clip-vit-base-patch32 | Vision | 21,704,784 | 934 |
| 12 | FacebookAI/xlm-roberta-base | Multilingual NLP | 20,698,741 | 827 |
| 13 | laion/clap-htsat-fused | Audio | 20,255,799 | 85 |
| 14 | FacebookAI/roberta-large | Classic NLP | 20,140,914 | 286 |
| 15 | Qwen/Qwen3-0.6B | Text-generation (LLM) | 18,989,268 | 1,241 |
Top 20 Text-Generation Models (LLMs Specifically)
| Rank | Model | Family | Downloads | Likes |
|---|---|---|---|---|
| 1 | Qwen/Qwen3-0.6B | Qwen | 18,989,268 | 1,241 |
| 2 | openai-community/gpt2 | GPT-2 (legacy) | 16,088,904 | 3,239 |
| 3 | Qwen/Qwen2.5-7B-Instruct | Qwen | 12,418,113 | 1,276 |
| 4 | Qwen/Qwen2.5-1.5B-Instruct | Qwen | 12,081,067 | 693 |
| 5 | Qwen/Qwen3-8B | Qwen | 11,735,972 | 1,087 |
| 6 | Qwen/Qwen3-4B-Instruct-2507 | Qwen | 10,991,777 | 841 |
| 7 | meta-llama/Llama-3.1-8B-Instruct | Llama | 9,751,974 | 5,825 |
| 8 | facebook/opt-125m | OPT (legacy) | 9,208,234 | 251 |
| 9 | Qwen/Qwen2.5-3B-Instruct | Qwen | 8,126,945 | 455 |
| 10 | meta-llama/Llama-3.2-1B-Instruct | Llama | 7,498,621 | 1,402 |
| 11 | openai/gpt-oss-20b | OpenAI (open release) | 7,304,172 | 4,604 |
| 12 | Qwen/Qwen3-32B | Qwen | 6,853,884 | 692 |
| 13 | Qwen/Qwen2.5-0.5B-Instruct | Qwen | 5,631,806 | 515 |
| 14 | openai/gpt-oss-120b | OpenAI (open release) | 4,566,280 | 4,772 |
| 15 | Qwen/Qwen3-4B | Qwen | 4,293,722 | 611 |
| 16 | deepseek-ai/DeepSeek-V3.2 | DeepSeek | 4,087,017 | 1,434 |
| 17 | deepseek-ai/DeepSeek-R1 | DeepSeek | 3,819,050 | 13,329 |
| 18 | Qwen/Qwen3-1.7B | Qwen | 3,535,359 | 465 |
| 19 | mistralai/Mistral-7B-Instruct-v0.2 | Mistral | 3,249,539 | 3,135 |
| 20 | meta-llama/Meta-Llama-3-8B | Llama | 3,135,302 | 6,532 |
Family Share of the Top 20 LLMs
| Family | Models in Top 20 | Combined Downloads |
|---|---|---|
| Qwen (Alibaba) | 11 | ~100,650,000 |
| Llama (Meta) | 3 | ~20,400,000 |
| GPT-OSS (OpenAI open release) | 2 | ~11,870,000 |
| DeepSeek | 2 | ~7,910,000 |
| Legacy (GPT-2, OPT-125m) | 2 | ~25,300,000 |
| Mistral | 1 | ~3,250,000 |
Seven Things the Rankings Tell You
- The single most-downloaded model on Hugging Face is not an LLM. sentence-transformers/all-MiniLM-L6-v2, an 80MB English-only embedding model, leads at 259M downloads. RAG and search infrastructure run on this one model more than on any LLM. The runner-up across all tasks (Qwen3-VL-2B-Instruct, multimodal) is the only model above 100M downloads besides MiniLM.
- Qwen owns the open-weight LLM ecosystem. 11 of the top 20 text-generation models on Hugging Face are Qwen variants. The combined ~100M downloads exceed every other family combined. Meta Llama (3 models, ~20M combined) is a distant second by deployment, despite being the dominant open-weight narrative in Western press.
- OpenAI's open release (gpt-oss-20b + gpt-oss-120b) is gaining ground. Two models, 11.9M combined downloads, 9,376 combined likes. The 120B variant in particular has a high like-to-download ratio (4,772 likes on 4.6M downloads), suggesting strong quality reception. OpenAI's first major open-weight release is being treated as a serious option, not a stunt.
- DeepSeek-R1 is the most-liked open LLM by a wide margin. 13,329 likes on 3.8M downloads is roughly 4x the like-rate of the typical top-20 model. R1 is the only open-weight reasoning model that consumers and researchers cite by name as a frontier-grade option; the like-to-download ratio confirms perception.
- BERT, GPT-2, and OPT are still in the top 20. google-bert/bert-base-uncased at 64M downloads, gpt2 at 16M, opt-125m at 9.2M. Six-to-eight year old models continue to do enormous volume because they are the default baselines in research code, tutorials, education, and lightweight fine-tuning pipelines.
- Chinese embeddings are competitive with Western embeddings. BAAI bge family (bge-small-en + bge-m3 + bge-large) collectively does ~83M downloads. sentence-transformers (all-MiniLM + paraphrase + all-mpnet) does ~343M. BAAI is the leading non-sentence-transformers embedding family on the registry.
- Vision is mostly CLIP. openai/clip-vit-large-patch14 (32M) and clip-vit-base-patch32 (22M) are the only vision-encoder models near the top of the all-task list, despite the rapid expansion of vision-language models like Qwen3-VL (which is classified as multimodal, not vision).
What This Means for AI Visibility
Open-weight model adoption shapes which models run inside agent stacks, on-device assistants, RAG retrievers, and custom fine-tunes. The Qwen-led ranking on Hugging Face suggests that any AI visibility programme assuming the Llama family dominates open-weight deployment is overestimating Meta and underestimating Alibaba. For Western brands optimising visibility across multi-language agents, the implication is that Chinese-language brand inclusion (in Qwen training corpora) matters more than visibility programmes typically weight it. For embeddings specifically, the dominance of all-MiniLM-L6-v2 and BAAI bge means that retrieval quality is largely a function of these few canonical embedding models, and content optimisation for retrieval should be tested against them rather than against frontier-vendor embedding APIs alone.
Methodology
Download and like counts pulled from the public Hugging Face Hub API on May 14, 2026 (the endpoints used: /api/models?sort=downloads&direction=-1&limit=30 for the all-task list and the same with filter=text-generation for the LLM list). Download counts are all-time, not monthly; this favours older models. Pipeline tags are HF's own classification (text-generation, sentence-similarity, image-text-to-text, etc.). Refreshed quarterly. Hugging Face download metrics include CI/CD pipeline pulls and mirror sync requests, so absolute numbers should be treated as proxy for cumulative deployment intensity rather than unique applications.
How Presenc AI Helps
Presenc AI monitors brand-mention rates across the major AI platforms whose downstream deployments draw from these open-weight base models. When a brand performs well on closed-frontier models but underperforms on agent stacks running Qwen or DeepSeek backends, the gap is traceable to training-data presence in the open-weight family. The Hugging Face rankings above are the input; brand-visibility-by-base-model is the output Presenc AI tracks for enterprise customers.