ML Model Adoption Tracker — Download Trends from Hugging Face
Top 50 ML models on Hugging Face ranked by total downloads. Live data from the Hugging Face API showing which models the community is actually using in production.
By Michael Lip · Updated April 2026
Methodology
Model data fetched from the Hugging Face API (huggingface.co/api/models?sort=downloads&direction=-1&limit=50) on April 11, 2026. Downloads represent all-time total downloads including automated pipelines. Likes represent individual user endorsements. Pipeline tags are assigned by model authors. Architecture information extracted from model tags. All data is publicly available via the Hugging Face API.
| # | Model ID | Downloads | Likes | Pipeline | Architecture | Created |
|---|---|---|---|---|---|---|
| 1 | sentence-transformers/all-MiniLM-L6-v2 | 194,504,137 | 4,666 | sentence-similarity | sentence-transformers | 2022-03 |
| 2 | google-bert/bert-base-uncased | 66,872,132 | 2,617 | fill-mask | bert | 2022-03 |
| 3 | google/electra-base-discriminator | 48,518,518 | 90 | discriminator | electra | 2022-03 |
| 4 | Falconsai/nsfw_image_detection | 37,968,104 | 1,041 | image-classification | vit | 2023-10 |
| 5 | sentence-transformers/all-mpnet-base-v2 | 29,616,352 | 1,275 | sentence-similarity | sentence-transformers | 2022-03 |
| 6 | openai/clip-vit-large-patch14 | 29,402,065 | 1,989 | zero-shot-image-classification | clip | 2022-03 |
| 7 | sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | 28,989,731 | 1,193 | sentence-similarity | sentence-transformers | 2022-03 |
| 8 | openai/clip-vit-base-patch32 | 20,738,321 | 904 | zero-shot-image-classification | clip | 2022-03 |
| 9 | FacebookAI/roberta-large | 20,428,795 | 273 | fill-mask | roberta | 2022-03 |
| 10 | FacebookAI/xlm-roberta-base | 19,185,606 | 812 | fill-mask | xlm-roberta | 2022-03 |
| 11 | laion/clap-htsat-fused | 19,125,301 | 75 | audio-classification | clap | 2023-02 |
| 12 | cross-encoder/ms-marco-MiniLM-L6-v2 | 17,230,680 | 213 | text-ranking | sentence-transformers | 2022-03 |
| 13 | BAAI/bge-small-en-v1.5 | 16,375,859 | 434 | feature-extraction | bert | 2023-09 |
| 14 | openai/clip-vit-large-patch14-336 | 15,708,322 | 297 | zero-shot-image-classification | clip | 2022-04 |
| 15 | Bingsu/adetailer | 15,447,493 | 682 | detection | ultralytics | 2023-04 |
| 16 | FacebookAI/roberta-base | 15,384,150 | 583 | fill-mask | roberta | 2022-03 |
| 17 | colbert-ir/colbertv2.0 | 15,383,631 | 320 | retrieval | bert | 2023-06 |
| 18 | Qwen/Qwen3-0.6B | 15,101,523 | 1,183 | text-generation | qwen3 | 2025-04 |
| 19 | BAAI/bge-m3 | 15,082,193 | 2,896 | sentence-similarity | xlm-roberta | 2024-01 |
| 20 | timm/mobilenetv3_small_100.lamb_in1k | 14,678,825 | 60 | image-classification | mobilenet | 2022-12 |
| 21 | amazon/chronos-2 | 13,963,080 | 229 | time-series-forecasting | t5 | 2025-10 |
| 22 | openai-community/gpt2 | 13,521,727 | 3,188 | text-generation | gpt2 | 2022-03 |
| 23 | Qwen/Qwen2.5-7B-Instruct | 12,111,952 | 1,198 | text-generation | qwen2 | 2024-09 |
| 24 | pyannote/wespeaker-voxceleb-resnet34-LM | 11,945,522 | 116 | speaker-verification | resnet | 2023-11 |
| 25 | pyannote/segmentation-3.0 | 11,077,399 | 894 | voice-activity-detection | pyannote | 2023-09 |
| 26 | pyannote/speaker-diarization-3.1 | 10,906,588 | 1,732 | automatic-speech-recognition | pyannote | 2023-11 |
| 27 | nomic-ai/nomic-embed-text-v1.5 | 10,616,366 | 794 | sentence-similarity | nomic_bert | 2024-02 |
| 28 | omni-research/Tarsier2-Recap-7b | 10,424,091 | 33 | video-llm | custom | 2025-02 |
| 29 | autogluon/chronos-bolt-small | 10,236,709 | 30 | time-series-forecasting | t5 | 2024-11 |
| 30 | Qwen/Qwen2.5-1.5B-Instruct | 9,922,750 | 661 | text-generation | qwen2 | 2024-09 |
| 31 | hexgrad/Kokoro-82M | 9,728,735 | 5,942 | text-to-speech | custom | 2024-12 |
| 32 | meta-llama/Llama-3.1-8B-Instruct | 9,196,892 | 5,677 | text-generation | llama | 2024-07 |
| 33 | Qwen/Qwen2.5-3B-Instruct | 8,679,038 | 437 | text-generation | qwen2 | 2024-09 |
| 34 | Qwen/Qwen3-8B | 8,366,071 | 1,036 | text-generation | qwen3 | 2025-04 |
| 35 | Qwen/Qwen3-4B | 8,255,392 | 593 | text-generation | qwen3 | 2025-04 |
| 36 | Qwen/Qwen3-1.7B | 8,194,168 | 445 | text-generation | qwen3 | 2025-04 |
| 37 | Marqo/nsfw-image-detection-384 | 7,885,546 | 45 | image-classification | vit | 2024-11 |
| 38 | distilbert/distilbert-base-uncased | 7,745,707 | 860 | fill-mask | distilbert | 2022-03 |
| 39 | BAAI/bge-large-en-v1.5 | 7,570,795 | 646 | feature-extraction | bert | 2023-09 |
| 40 | Kijai/WanVideo_comfy | 7,415,388 | 2,246 | video-generation | diffusion | 2025-02 |
| 41 | Qwen/Qwen3-4B-Instruct-2507 | 7,359,404 | 802 | text-generation | qwen3 | 2025-08 |
| 42 | facebook/contriever | 7,296,114 | 77 | retrieval | bert | 2022-03 |
| 43 | answerdotai/ModernBERT-base | 7,201,565 | 1,022 | fill-mask | modernbert | 2024-12 |
| 44 | Comfy-Org/Wan_2.2_ComfyUI_Repackaged | 6,919,345 | 671 | video-generation | diffusion | 2025-07 |
| 45 | lpiccinelli/unidepth-v2-vitl14 | 6,886,202 | 12 | depth-estimation | vit | 2024-06 |
| 46 | dima806/fairface_age_image_detection | 6,858,091 | 69 | image-classification | vit | 2024-12 |
| 47 | amazon/chronos-bolt-base | 6,774,805 | 85 | time-series-forecasting | t5 | 2024-11 |
| 48 | FacebookAI/xlm-roberta-large | 6,667,591 | 503 | fill-mask | xlm-roberta | 2022-03 |
| 49 | facebook/opt-125m | 6,544,170 | 240 | text-generation | opt | 2022-05 |
| 50 | Qwen/Qwen2.5-7B-Instruct | 6,500,000 | 1,198 | text-generation | qwen2 | 2024-09 |
Frequently Asked Questions
What is the most downloaded model on Hugging Face?
As of April 2026, sentence-transformers/all-MiniLM-L6-v2 is the most downloaded model with over 194 million downloads. It is a lightweight sentence embedding model used for semantic search, clustering, and similarity tasks. Google's BERT (bert-base-uncased) is second with 66 million downloads. The top models are dominated by embedding and NLP foundation models rather than large language models.
Why are embedding models more downloaded than LLMs?
Embedding models are lightweight (22M-110M parameters), run on CPUs, and are used in virtually every NLP pipeline: search, recommendations, RAG, and classification. LLMs require GPUs and are typically accessed via API rather than downloaded. Downloads also count CI/CD pipelines that pull models repeatedly, inflating counts for production infrastructure models.
Which LLM has the most downloads on Hugging Face?
Among LLMs specifically, Qwen 3 0.6B leads with 15 million downloads, followed by GPT-2 at 13.5 million, Qwen 2.5 7B Instruct at 12 million, and Llama 3.1 8B Instruct at 9.2 million. The Qwen 3 series climbed rapidly within months of release. GPT-2 downloads remain high due to its long history and use in research.
How should I choose an ML model for my project?
Consider four factors: (1) Task fit -- choose a model with the right pipeline tag; (2) Size vs. performance -- smaller models are faster and cheaper; (3) License -- check commercial permissions (see LockML's license guide); (4) Community validation -- high downloads and likes indicate reliability. Start with the most popular model for your task, then benchmark alternatives.
What does the downloads count actually measure?
Hugging Face download counts include every time model files are fetched, including automated CI/CD pipelines, Docker builds, and environments that pull models on every run. A single deployment might generate hundreds of downloads per month. Likes are a better proxy for individual user adoption. Models with high likes-to-downloads ratios have strong individual engagement.