Best Hugging Face Models: Ranked on Real Download Data

TL;DR: If you want one answer, the most-used Hugging Face model on the Hub is sentence-transformers/all-MiniLM-L6-v2 at 254,324,263 downloads in 30 days, which is 29.8% of every download across this list, and it is a small embedding model rather than a chat model. If you want the answer for your job, the picks table is directly below. This ranks 300 Hugging Face models on what the Hub API reports, not on opinion.

Last updated: 11 September 2026. Data pulled: 11 September 2026, from the free public Hugging Face Hub API. Models ranked: 300. Downloads in the last 30 days: 853,841,158. All-time downloads across the list: 8,382,324,609.

Every “best Hugging Face models” article I can find ranks models by vibes, or by whichever LLM was trending the week it was written. This one ranks them by what the Hub itself reports: downloads, likes, licence, and how recently each model was touched.

I have spent 15 years in software and SEO and tested more than 500 AI tools, and the reason I keep coming back to download data is that it is the only honest signal of production use. Benchmarks tell you what a model can do.

Downloads tell you what people are actually shipping. On Hugging Face those two answers are further apart than almost anyone expects.

Best Hugging Face models by job: the short answer

Nobody needs “the best model”. They need an embedding model, or something that transcribes audio, or a chat model small enough to run on the laptop they own. So here is one pick per job before any of the analysis.

The job	Pick	30-day downloads	Licence	Quality score	What it is
Semantic search and RAG	nomic-ai/nomic-embed-text-v1.5	16,124,079	apache-2.0	87/100	turns text into vectors you can search
Re-ranking search results	nvidia/llama-nemotron-rerank-1b-v2	1,097,075	other	94/100	reorders the hits retrieval already found
Searching scanned documents	vidore/colqwen2.5-v0.2	328,936	mit	84/100	retrieves from page images, no OCR step
Chat you host yourself	farbodtavakkoli/OTel-2.0-LLM-31B-IT	7,164,611	apache-2.0	97/100	open weights you download rather than call
Reading images and documents	Qwen/Qwen3.6-35B-A3B-FP8	12,066,045	apache-2.0	92/100	vision-language, the busiest group here
Transcribing audio	argmaxinc/whisperkit-coreml	11,634,364	mit	100/100	speech to text
Generating speech	pnnbao-ump/VieNeu-TTS-v3-Turbo	529,723	apache-2.0	92/100	text to speech
Running on a laptop	GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-GGUF	374,141	apache-2.0	88/100	generative and under 4B parameters
Forecasting a time series	autogluon/chronos-2	9,038,338	apache-2.0	80/100	numbers over time, not language

How each pick is chosen, so you can disagree with it. For every job I take the five most-downloaded models the Hub tags for that task, then pick whichever of those five scores highest on the quality score explained further down. Adoption first, quality as the tiebreak.

That rule is deliberate: scoring alone put a 661,992-download embedding model ahead of the 254,324,263-download one, which would contradict the argument this whole report rests on. Nothing here is hand-picked, nobody paid for a slot, and there are no affiliate links, because Hugging Face models are free.

Two honest caveats about that table. The Hub’s own task tag decides which group a model lands in, and those tags are imperfect: a good number of chat models are filed under image-text-to-text because they also accept images. And a model has to clear an adoption floor before it is ranked here at all, so a genuinely excellent model released three weeks ago will not appear.

The numbers worth quoting

• 300 Hugging Face models ranked as of 11 September 2026, with 853,841,158 downloads in 30 days and 8.38B all-time, from 111 publishers.
• sentence-transformers/all-MiniLM-L6-v2 alone is 29.8% of 30-day downloads (254,324,263). It is a sentence-similarity model, not a chat model.
• Embeddings & Retrieval has only 18 models but takes 47.6% of downloads. Vision & Multimodal has 121 models and takes 21.7%.
• That is 22.6M downloads per model versus 1.5M, roughly a 15x difference.
• The top ten models take 53.9% of all downloads. The remaining 290 share 46.1%.
• 78 of the 116 chat models whose name states a size sit in the 20B to 40B band (67%), against one at 10B to 20B.
• 142 of 300 models (47%) are quantized or format-converted re-uploads of someone else’s weights, accounting for 22.1% of downloads.
• 75% are Apache-2.0 or MIT. 9 state no licence at all, four are non-commercial, and 0 is access-gated.
• Likes measure attention, not use. The most-liked model, Qwen/Qwen3.8-27B at 14,669 likes, ranks only #18 by downloads.
• Maintenance is strong: median 28 days since the last update, 159 models updated within 30 days, and only 0 untouched for over a year.

Every figure is reproducible from the tables below. Method and citation line at the bottom.

Which kinds of model are actually downloaded most?

Embeddings & Retrieval models. Not chat models, not image generators. 18 embedding and retrieval models take 47.6% of every download in this list, while 121 Vision & Multimodal models take 21.7%.

Task group	Models	30-day downloads	Share of downloads	Downloads per model
Embeddings & Retrieval	18	406,305,820	47.6%	22.6M
Vision & Multimodal	121	185,258,344	21.7%	1.5M
Other	60	112,472,701	13.2%	1.9M
Text Generation & Chat	69	100,052,774	11.7%	1.5M
Speech & Audio	22	33,825,652	4.0%	1.5M
Tabular, Time-Series & Robotics	2	9,522,125	1.1%	4.8M
Text Classification	6	5,286,879	0.6%	881.1k
Image Generation	2	1,116,863	0.1%	558.4k
All 8 groups	300	853,841,158	100%	2.8M

This is the single most useful thing in the dataset, and it is almost the exact inverse of the public conversation.

The reason is structural. An embedding model is infrastructure: if you run semantic search, a RAG pipeline, or a recommendation system, you call it on every document and every query, and you pull the weights into every container you deploy. A chat model is a destination: you download it once, or you never download it at all because you call somebody’s API instead.

So download counts on Hugging Face measure something specific. They measure how often a model gets pulled into a build, which correlates with infrastructure use rather than user-facing excitement. That is worth knowing before you cite a download number as evidence that a model is “the best”.

The practical read for a buyer: if you are choosing an embedding model, this data is extremely informative, because the whole market is here and heavily used. If you are choosing a chat model, treat downloads as one weak signal among several.

The 10 best Hugging Face models by quality score

Rank	Model	Task	30-day downloads	Likes	Licence	Quality score
1	argmaxinc/whisperkit-coreml	automatic-speech-recognition	11,634,364	206	mit	100/100
2	answerdotai/answerai-colbert-small-v1	other	544,075	163	apache-2.0	100/100
3	Comfy-Org/MiniMax-H3	other	18,813,055	1,761	other	99/100
4	Comfy-Org/z_image_turbo	other	7,444,224	874	apache-2.0	99/100
5	Comfy-Org/Qwen-Image_ComfyUI	other	2,483,092	483	apache-2.0	98/100
6	moonshotai/Kimi-K3	image-text-to-text	2,319,885	11,275	other	98/100
7	farbodtavakkoli/OTel-2.0-LLM-31B-IT	text-generation	7,164,611	15	apache-2.0	97/100
8	Comfy-Org/Qwen-Image-Edit_ComfyUI	other	1,668,219	472	apache-2.0	97/100
9	kingabzpro/wav2vec2-large-xls-r-300m-Urdu	automatic-speech-recognition	1,667,280	14	apache-2.0	95/100
10	circlestone-labs/Anima	other	1,009,755	2,210	other	95/100

This ranking weighs adoption alongside growth, maintenance and trust, so it surfaces models that are both used and well kept rather than only enormous. argmaxinc/whisperkit-coreml leads at 100/100 on 11,634,364 downloads. Note what happens further down the table: several entries have download counts far below the raw leaders and still place highly, because they are growing fast, freshly updated, properly licensed and not gated.

How the quality score is calculated

Four weighted pillars: adoption 40%, maintenance 25%, growth 20%, trust 15%.

• Adoption (40%) blends all-time downloads and likes, log-scaled so a handful of enormous outliers do not flatten everything below them.
• Maintenance (25%) is how recently the model was updated, decaying the longer it sits untouched. The list does well here: the median model was updated 28 days ago, 159 of 300 (53%) within 30 days, 265 within 90, and only 0 has gone more than a year. Read “updated” carefully though, because the Hub’s last-modified date moves on any repository change, a README edit included. It says somebody is still paying attention to the repo, not that the model was retrained.
• Growth (20%) is the 30-day change in downloads, measured against our own snapshot from roughly a month earlier. It is blank for models too new to have history, and those are scored neutrally rather than penalised.
• Trust (15%) is structural rather than subjective: does the model state a licence, is its author identified, and is it free of an access gate. It is the pillar that most often separates an otherwise-identical pair.

Because every model here already cleared an adoption floor, the scores cluster tightly: 79 to 100, median 83. Use the score to order models inside a task group.

It is not a verdict on whether a model is any good, and I would not read a 83 as a warning. Hugging Face also exposes an internal “trending score”, and 219 of the 300 models carry a non-zero one, but its scale is not publicly documented so it is not used in the ranking at all.

The 10 Most-Downloaded Hugging Face Models

Rank	Model	Task	30-day downloads	Share	Quality score
1	sentence-transformers/all-MiniLM-L6-v2	sentence-similarity	254,324,263	29.8%	85/100
2	cross-encoder/ms-marco-MiniLM-L6-v2	text-ranking	87,478,105	10.2%	90/100
3	Comfy-Org/MiniMax-H3	other	18,813,055	2.2%	99/100
4	trl-internal-testing/tiny-Qwen2ForCausalLM-2.5	text-generation	18,620,166	2.2%	84/100
5	Comfy-Org/stable-diffusion-v1-5-archive	other	17,379,879	2.0%	90/100
6	nomic-ai/nomic-embed-text-v1.5	sentence-similarity	16,124,079	1.9%	87/100
7	intfloat/multilingual-e5-small	sentence-similarity	12,369,093	1.4%	80/100
8	Qwen/Qwen3.6-35B-A3B-FP8	image-text-to-text	12,066,045	1.4%	92/100
9	argmaxinc/whisperkit-coreml	automatic-speech-recognition	11,634,364	1.4%	100/100
10	unsloth/Qwen3.8-27B-GGUF	other	11,127,203	1.3%	87/100

Compare this against the quality table. They share almost nothing, and that is the point. sentence-transformers/all-MiniLM-L6-v2 is first here and does not lead on quality score; argmaxinc/whisperkit-coreml leads on quality score and is not first here.

Neither ranking is wrong. Downloads tell you what is battle-tested and safe to standardise on, while the quality score tells you what is well maintained and moving.

For infrastructure you will keep for years, weight downloads more heavily. For something you are adopting now, weight the score.

Downloads are concentrated to an extreme degree

sentence-transformers/all-MiniLM-L6-v2 alone accounts for 29.8% of 30-day downloads. The top three take 42.2%, the top ten 53.9%, and the top fifty 77.5%.

The median model in this list gets 697,663 downloads in 30 days, so the gap between the leader and the middle is roughly four orders of magnitude. What makes the leader interesting is how old and how small it is. `all-MiniLM-L6-v2` is a compact sentence-transformer that has been the default embedding model in countless tutorials, frameworks and starter templates for years. Its dominance is a lesson in defaults: being the thing that gets copy-pasted into every quickstart compounds far faster than being the best model in a benchmark table.

Likes tell you what is exciting, not what is deployed

Model	Likes	30-day downloads	Rank by downloads	Task
Qwen/Qwen3.8-27B	14,669	7,322,476	#18	image-text-to-text
moonshotai/Kimi-K3	11,275	2,319,885	#59	image-text-to-text
sentence-transformers/all-MiniLM-L6-v2	5,773	254,324,263	#1	sentence-similarity
MiniMaxAI/MiniMax-H3	5,133	5,080,204	#26	image-text-to-video
Qwen/Qwen3.8-Flash-Next	5,085	564,079	#182	image-text-to-text
bigscience/bloom	5,043	0	#285	text-generation
baidu/Unlimited-OCR	4,212	2,724,383	#48	image-text-to-text
deepseek-ai/DeepSeek-V4-Flash-0731	3,928	4,393,881	#28	text-generation
unsloth/Qwen3.8-27B-GGUF	3,855	11,127,203	#10	other
google/gemma-4-31B-it	3,757	8,744,810	#14	image-text-to-text

Qwen/Qwen3.8-27B is the most-liked model at 14,669 likes, and ranks #18 by downloads with 7,322,476 in 30 days. The list carries 179,513 likes in total and they cluster on frontier-scale releases that most people admire and few people can actually run.

That is not a criticism of likes. They answer a different question, and both columns are in the tables so you can read either.

Best Hugging Face models by task

Each group below is the top six by quality score within that task, with downloads, likes and licence alongside.

Best embedding and retrieval models

These turn text into vectors for semantic search, RAG pipelines, and re-ranking. They are small, unglamorous, and by download volume they are the most used models on the entire Hub.

18 models in this group, 406,305,820 downloads in the last 30 days (47.6% of the total).

Model	30-day downloads	Likes	Licence	Quality score
nvidia/llama-nemotron-rerank-1b-v2	1,097,075	64	other	94/100
nvidia/Nemotron-3-Embed-1B-BF16	661,992	147	other	92/100
sentence-transformers/all-MiniLM-L12-v2	4,093,501	328	apache-2.0	91/100
cross-encoder/ms-marco-MiniLM-L6-v2	87,478,105	316	apache-2.0	90/100
zeroentropy/zerank-2-reranker	644,922	117	apache-2.0	90/100
nomic-ai/nomic-embed-text-v1.5	16,124,079	908	apache-2.0	87/100

Best vision and multimodal models

Vision-language models that handle image understanding, visual question answering, OCR, and multimodal chat in a single model. This is the most crowded category on the list.

121 models in this group, 185,258,344 downloads in the last 30 days (21.7% of the total).

Model	30-day downloads	Likes	Licence	Quality score
moonshotai/Kimi-K3	2,319,885	11,275	other	98/100
tencent/HunyuanOCR	678,306	817	other	95/100
Qwen/Qwen3.6-35B-A3B-FP8	12,066,045	381	apache-2.0	92/100
DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF	1,032,720	607	apache-2.0	92/100
unsloth/gemma-4-E2B-it-GGUF	594,026	306	apache-2.0	92/100
Comfy-Org/Wan_2.2_ComfyUI_Repackaged	5,298,475	870	apache-2.0	91/100

Other notable models

Models whose Hugging Face task tag does not map cleanly onto the groups above, including reinforcement learning, depth estimation, and unclassified releases.

60 models in this group, 112,472,701 downloads in the last 30 days (13.2% of the total).

Model	30-day downloads	Likes	Licence	Quality score
answerdotai/answerai-colbert-small-v1	544,075	163	apache-2.0	100/100
Comfy-Org/MiniMax-H3	18,813,055	1,761	other	99/100
Comfy-Org/z_image_turbo	7,444,224	874	apache-2.0	99/100
Comfy-Org/Qwen-Image_ComfyUI	2,483,092	483	apache-2.0	98/100
Comfy-Org/Qwen-Image-Edit_ComfyUI	1,668,219	472	apache-2.0	97/100
circlestone-labs/Anima	1,009,755	2,210	other	95/100

Best text generation and chat models

The models behind most chat assistants and text-completion tools built on Hugging Face, including the open-weight releases people actually deploy rather than only benchmark.

69 models in this group, 100,052,774 downloads in the last 30 days (11.7% of the total).

Model	30-day downloads	Likes	Licence	Quality score
farbodtavakkoli/OTel-2.0-LLM-31B-IT	7,164,611	15	apache-2.0	97/100
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	1,237,895	424	other	94/100
deepseek-ai/DeepSeek-V4-Flash-0731	4,393,881	3,928	mit	93/100
cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit	948,375	59	apache-2.0	92/100
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8	640,081	357	other	92/100
swiss-ai/Apertus-8B-Instruct-2509	508,878	491	apache-2.0	92/100

Best speech and audio models

Transcription, text-to-speech, and audio classification. Speech recognition in particular has quietly become one of the most reliably downloaded categories on the Hub.

22 models in this group, 33,825,652 downloads in the last 30 days (4.0% of the total).

Model	30-day downloads	Likes	Licence	Quality score
argmaxinc/whisperkit-coreml	11,634,364	206	mit	100/100
kingabzpro/wav2vec2-large-xls-r-300m-Urdu	1,667,280	14	apache-2.0	95/100
pnnbao-ump/VieNeu-TTS-v3-Turbo	529,723	60	apache-2.0	92/100
nvidia/parakeet-tdt-0.6b-v3	702,506	1,115	cc-by-4.0	91/100
MahmoudAshraf/mms-300m-1130-forced-aligner	2,777,141	103	cc-by-nc-4.0	90/100
k2-fsa/OmniVoice	1,194,599	1,369	none	88/100

Best tabular, time-series and robotics models

An emerging mix of forecasting and robotics foundation models. Tiny by model count and punching far above its weight on downloads.

2 models in this group, 9,522,125 downloads in the last 30 days (1.1% of the total).

Model	30-day downloads	Likes	Licence	Quality score
autogluon/chronos-2	9,038,338	49	apache-2.0	80/100
google/timesfm-3.0-pytorch	483,787	718	other	80/100

Best text classification models

Sentiment, masked-token prediction, named-entity tagging, and safety classification. Old-fashioned by 2026 standards and still doing enormous amounts of production work.

6 models in this group, 5,286,879 downloads in the last 30 days (0.6% of the total).

Model	30-day downloads	Likes	Licence	Quality score
protectai/deberta-v3-base-prompt-injection-v2	837,634	115	apache-2.0	93/100
protectai/unbiased-toxic-roberta-onnx	271,155	7	apache-2.0	90/100
biohub/ESMC-6B	2,588,347	31	mit	88/100
protectai/xlm-roberta-base-language-detection-onnx	191,658	6	mit	88/100
cointegrated/rubert-tiny-toxicity	439,907	48	mit	80/100
LocalAI-io/privacy-filter-nemotron-GGUF	958,178	0	apache-2.0	79/100

Best image generation models

Text-to-image generation. A small category here by count, because most image-generation traffic sits with a handful of well-known checkpoints and their community fine-tunes.

2 models in this group, 1,116,863 downloads in the last 30 days (0.1% of the total).

Model	30-day downloads	Likes	Licence	Quality score
nphSi/Z-Image-Lora	265,540	149	apache-2.0	92/100
RunDiffusion/Juggernaut-XL-v9	851,323	435	creativeml-openrail-m	88/100

Best open-source LLMs on Hugging Face, by what you can actually run

The question people ask is almost never “which large language model is best”. It is “which one is best that fits the hardware I have”, phrased as a size: a 12B, something under 10GB, something that will run on two consumer cards.

So here is the chat side of the list banded by parameter count. This covers the 155 generative models on the list, 116 of which (75%) state a size in their name, which is the only place the Hub API exposes one. Where an id names both a total and an active-expert count, the total is used, because total parameters is what has to fit.

Parameter band	Models	30-day downloads	Share of all downloads	Most downloaded in the band
Under 4B	8	6,973,374	0.8%	Qwen/Qwen3.5-2B
4B to 10B	20	25,062,793	2.9%	Qwen/Qwen3.5-4B
10B to 20B	1	646,541	0.1%	yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF
20B to 40B	78	150,438,403	17.6%	Qwen/Qwen3.6-35B-A3B-FP8
40B and above	9	6,410,171	0.8%	Qwen/Qwen3.5-122B-A10B-FP8

The middle of the range has emptied out, and that is the finding here. 78 of the 116 sized chat models (67%) sit in the 20B to 40B band, taking 17.6% of all downloads on this list, led by Qwen/Qwen3.6-35B-A3B-FP8. The 10B to 20B band holds exactly one model.

That is not an artefact of restricting this table to chat models either: across all 300 ranked models, sized or not, only 6 land between 10B and 20B. If you came here looking for the best 12B or 14B model, the honest answer is that the field largely stopped building them.

The reason is mixture-of-experts. 34 models here are MoE builds, 7.2% of downloads, and the naming convention gives the game away: Qwen/Qwen3.6-35B-A3B-FP8 states a total and a much smaller active-parameter count.

A model like that loads at its full size but only computes with a fraction of itself per token, so it runs at roughly small-model speed with large-model quality. Once that worked, there was little reason to train a dense 13B, and the people who would have downloaded one now download a 30-something-B MoE instead.

What that means for your hardware, roughly. A 4-bit GGUF build needs about 0.6GB of memory per billion parameters plus room for context, so a 30B MoE at 4-bit lands near 20GB, which is a 24GB card or an Apple Silicon machine with enough unified memory.

90 models on the full list (11.2% of downloads) ship as GGUF specifically so `llama.cpp` and the tools built on it can load them that way. Treat that arithmetic as a planning figure and check the file sizes on the model card, since quantization level and context length both move it.

At the extremes: 28 chat models are under 10B and take 3.8% of downloads, Qwen/Qwen3.5-4B being the most pulled. That is the band to look in if you want something that runs on a laptop with no discrete GPU.

Nine are 40B and above, taking 0.8%, and Qwen/Qwen3.5-122B-A10B-FP8 leads them. That last figure is the one worth sitting with: the largest open models get admired far more than they get downloaded, because almost nobody has the hardware.

Best embedding models for semantic search, RAG and re-ranking

This is the category that pays the rent on the Hub, and the one where the download data is most worth trusting, because the entire market is here and every model in it is used heavily.

Model	What it does	30-day downloads	Licence	Quality score
sentence-transformers/all-MiniLM-L6-v2	Text embeddings	254,324,263	apache-2.0	85/100
cross-encoder/ms-marco-MiniLM-L6-v2	Re-ranking	87,478,105	apache-2.0	90/100
nomic-ai/nomic-embed-text-v1.5	Text embeddings	16,124,079	apache-2.0	87/100
intfloat/multilingual-e5-small	Text embeddings	12,369,093	mit	80/100
Qwen/Qwen3-Embedding-0.6B	Text embeddings	7,806,497	apache-2.0	81/100
intfloat/multilingual-e5-base	Text embeddings	7,189,365	mit	86/100
ibm-granite/granite-embedding-small-english-r2	Text embeddings	6,324,429	apache-2.0	84/100
sentence-transformers/all-MiniLM-L12-v2	Text embeddings	4,093,501	apache-2.0	91/100
Qwen/Qwen3-Reranker-4B	Re-ranking	2,546,793	apache-2.0	79/100
Qwen/Qwen3-VL-Reranker-2B	Re-ranking	1,676,336	apache-2.0	87/100
Qwen/Qwen3-VL-Embedding-2B	Text embeddings	1,307,966	apache-2.0	80/100
nvidia/llama-nemotron-rerank-1b-v2	Re-ranking	1,097,075	other	94/100

22 models here do retrieval work of some kind, 47.7% of all downloads. They split into three jobs that people routinely confuse:

• Embedding models turn a document or a query into a vector. You run one over your whole corpus once, then over every query forever, which is why the download counts are what they are.
• Re-rankers (6 models, 11.0% of downloads) take the shortlist an embedding search returned and reorder it properly. They cost far more per pair at inference time and are far more accurate, so you run one over the top 50 hits rather than the whole corpus. cross-encoder/ms-marco-MiniLM-L6-v2 at 87,478,105 downloads is the workhorse.
• Document-image retrieval (4 models, the ColPali and ColQwen family, led by vidore/colqwen2-v1.0) skips OCR entirely and embeds the page image. If your corpus is scanned PDFs, invoices or slides, this is a different and usually better answer than running OCR and then embedding the text.

Picking one comes down to four things, in this order. Does it fit your language? 6 of these are explicitly multilingual, intfloat/multilingual-e5-small being the most used, and an English-only model on non-English text fails quietly rather than loudly. How big is the vector? Larger dimensions cost storage and query time in your vector database forever, so a 384-dimension model that is nearly as accurate is often the better engineering call. Can you afford it at your volume? These run on every document you index. And does the licence work? Most of this group is Apache-2.0 or MIT, but not all of it, and a non-commercial embedding model quietly poisons a commercial product.

For accuracy comparisons specifically, download counts are the wrong tool and the MTEB leaderboard is the right one. It benchmarks retrieval quality across dozens of tasks.

Read it alongside this table rather than instead of it: MTEB tells you which model scores best, this tells you which one the rest of the industry has already bet on, and those are genuinely different questions. Then test both on your own documents, because domain fit beats leaderboard position more often than anyone likes to admit.

What changed on the hub in the last 30 days

deepseek-ai/DeepSeek-V4-Flash-0731 grew 525% over the last 30 days, followed by Comfy-Org/MiniMax-H3 at 499%. Of the 150 models with enough history to measure, 107 grew and 43 declined.

Model	30-day trend	30-day downloads	Task	Quality score
deepseek-ai/DeepSeek-V4-Flash-0731	+525%	4,393,881	text-generation	93/100
Comfy-Org/MiniMax-H3	+499%	18,813,055	other	99/100
unsloth/gemma-4-12B-it-qat-GGUF	+426%	1,336,575	any-to-any	91/100
RunDiffusion/Juggernaut-XL-v9	+376%	851,323	text-to-image	88/100
google/gemma-4-E4B-it-qat-q4_0-gguf	+219%	652,126	any-to-any	89/100
DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF	+202%	1,032,720	image-text-to-text	92/100
iitolstykh/mivolo_v2	+193%	2,647,955	other	85/100
Qwen/Qwen3-VL-Reranker-2B	+189%	1,676,336	text-ranking	87/100
protectai/deberta-v3-base-prompt-injection-v2	+185%	837,634	text-classification	93/100
unsloth/inkling-GGUF	+163%	665,869	image-text-to-text	89/100
google/gemma-4-12B-it-qat-q4_0-gguf	+158%	780,934	any-to-any	90/100
nvidia/parakeet-tdt-0.6b-v3	+151%	702,506	automatic-speech-recognition	91/100

And the other direction:

Model	30-day trend	30-day downloads	Task	Quality score
Lightricks/LTX-2.3	-33%	1,194,236	image-to-video	79/100
cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit	-32%	2,256,988	image-text-to-text	80/100
google/gemma-4-31B-it	-21%	8,744,810	image-text-to-text	83/100
google/gemma-4-E2B-it	-20%	3,188,731	any-to-any	81/100
google/gemma-4-26B-A4B-it	-19%	8,867,625	image-text-to-text	84/100
unsloth/Qwen3.6-35B-A3B-NVFP4	-18%	1,640,310	image-text-to-text	79/100

Lightricks/LTX-2.3 is down 33%, the steepest decline here. Most decliners are the previous generation of the same model family, which is what healthy replacement looks like in download data rather than a sign of a problem.

Most of the movement is repackaging, not new models

142 of the 300 ranked models (47%) are quantized or converted re-uploads in formats like GGUF, AWQ, NVFP4 and INT4, together 22.1% of downloads. Whole publishers on this list exist to do nothing else. It tells you where the real bottleneck is: not capability, but getting existing capability onto hardware people can afford.

That layer is worth understanding rather than dismissing, because it is where a lot of the practical value is. Somebody takes a 27B model nobody can run on a home machine, quantizes it to 4-bit, tests that it did not break, and publishes it.

That is real work. It also means the model card you are reading is often not written by the people who trained the weights, so check what the original licence was before you assume the re-upload inherited it correctly.

Are there uncensored or NSFW models on Hugging Face?

Yes, openly, and this dataset can tell you how many. Sixteen of the 300 ranked models advertise removed refusal behaviour in their name, using words like uncensored, abliterated or heretic, from 12 publishers. Together they take 16,807,736 downloads in 30 days, 2.0% of the list.

Model	Publisher	30-day downloads	Base model it modifies	Licence
JonathanColetti/Qwen3.8-27B-Uncensored-GGUF	JonathanColetti	2,812,743	Qwen3.8-27B	apache-2.0
huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF	huihui-ai	2,441,040	Huihui-Qwen3.8-27B	apache-2.0
HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF	HauhauCS	1,908,917	Qwen3.8-27B	apache-2.0
0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF	0bserverx	1,721,344	Qwen3.8-27B	apache-2.0
OBLITERATUS/Qwen3.8-27B-OBLITERATED	OBLITERATUS	1,133,995	Qwen3.8-27B	apache-2.0
DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF	DavidAU	1,032,720	Qwen3.5-9B-The-Defiant-Fable	apache-2.0
Bahushruth/Qwen3.6-35B-A3B-abliterated-v4	Bahushruth	935,126	Qwen3.6-35B-A3B	apache-2.0
HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive	HauhauCS	858,482	Qwen3.5-9B	apache-2.0

Three things that table shows. Almost all of them are the same handful of base models, re-tuned; 69% are also quantized re-uploads, so they are packaging and modification stacked on top of each other.

They inherit the base model’s licence, usually Apache-2.0, which permits the modification but says nothing about what you then do with the output. And “abliterated” describes a specific technique, identifying the direction in the model’s activations that corresponds to refusal and suppressing it, which reliably removes the refusals and also degrades the model’s judgement in ways that are hard to measure. If you are evaluating one for a real product, benchmark it against its base model rather than assuming the only thing that changed is the guardrail.

Are Hugging Face models free to use? Licences in practice

Mostly yes, and more permissively than people assume. 188 models are Apache-2.0 and 37 are MIT, so 75% of the list carries a licence that allows commercial use with minimal conditions.

License	Models	Share
apache-2.0	188	63%
other	49	16%
mit	37	12%
none stated	9	3%
cc-by-4.0	5	2%
creativeml-openrail-m	2	1%
cc-by-nc-4.0	2	1%
openmdw-1.1	2	1%

The exceptions are where the care is needed. Nine models state no licence at all, which legally is the most restrictive outcome rather than the least, because absent a grant you have no permission. four carry a non-commercial licence, and 0 is access-gated, meaning you must accept terms before downloading.

A caveat on the “other” bucket: Hugging Face reports it for custom licences, which includes several widely used model families whose terms are real but non-standard. If a model matters to your product, read its actual licence file rather than trusting the tag.

The full list: All 300 Hugging Face models ranked

Every ranked model, ordered by quality score, with the inputs shown so you can check the arithmetic. Downloads are as reported by the Hub API. Click any model to open its official Hugging Face page and model card.

#	Model	Publisher	Task	30-day downloads	All-time	Likes	Updated	Licence	Quality
1	argmaxinc/whisperkit-coreml	argmaxinc	automatic-speech-recognition	11,634,364	71,701,537	206	2026-08-19	mit	100
2	answerdotai/answerai-colbert-small-v1	answerdotai	other	544,075	55,792,959	163	2026-08-17	apache-2.0	100
3	Comfy-Org/MiniMax-H3	Comfy-Org	other	18,813,055	29,330,247	1,761	2026-09-06	other	99
4	Comfy-Org/z_image_turbo	Comfy-Org	other	7,444,224	36,948,641	874	2026-08-17	apache-2.0	99
5	Comfy-Org/Qwen-Image_ComfyUI	Comfy-Org	other	2,483,092	24,821,139	483	2026-08-17	apache-2.0	98
6	moonshotai/Kimi-K3	moonshotai	image-text-to-text	2,319,885	4,187,772	11,275	2026-09-02	other	98
7	farbodtavakkoli/OTel-2.0-LLM-31B-IT	farbodtavakkoli	text-generation	7,164,611	11,925,847	15	2026-09-08	apache-2.0	97
8	Comfy-Org/Qwen-Image-Edit_ComfyUI	Comfy-Org	other	1,668,219	11,315,539	472	2026-08-17	apache-2.0	97
9	kingabzpro/wav2vec2-large-xls-r-300m-Urdu	kingabzpro	automatic-speech-recognition	1,667,280	15,511,649	14	2026-06-24	apache-2.0	95
10	circlestone-labs/Anima	circlestone-labs	other	1,009,755	4,738,758	2,210	2026-08-24	other	95
11	tencent/HunyuanOCR	tencent	image-text-to-text	678,306	5,832,422	817	2026-08-28	other	95
12	nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	nvidia	text-generation	1,237,895	5,590,124	424	2026-08-25	other	94
13	nvidia/llama-nemotron-rerank-1b-v2	nvidia	text-ranking	1,097,075	3,365,212	64	2026-08-26	other	94
14	kernels-community/flash-attn3	kernels-community	other	643,060	3,221,198	49	2026-09-01	bsd-3-clause	94
15	deepseek-ai/DeepSeek-V4-Flash-0731	deepseek-ai	text-generation	4,393,881	5,834,974	3,928	2026-08-01	mit	93
16	protectai/deberta-v3-base-prompt-injection-v2	protectai	text-classification	837,634	6,867,109	115	2026-07-09	apache-2.0	93
17	Qwen/Qwen3.6-35B-A3B-FP8	Qwen	image-text-to-text	12,066,045	40,808,526	381	2026-04-24	apache-2.0	92
18	DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF	DavidAU	image-text-to-text	1,032,720	1,543,720	607	2026-08-24	apache-2.0	92
19	cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit	cyankiwi	text-generation	948,375	3,759,853	59	2026-07-21	apache-2.0	92
20	nvidia/Nemotron-3-Embed-1B-BF16	nvidia	sentence-similarity	661,992	1,145,965	147	2026-08-27	other	92
21	nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8	nvidia	text-generation	640,081	7,323,301	357	2026-08-24	other	92
22	unsloth/gemma-4-E2B-it-GGUF	unsloth	image-text-to-text	594,026	4,368,393	306	2026-07-17	apache-2.0	92
23	pnnbao-ump/VieNeu-TTS-v3-Turbo	pnnbao-ump	text-to-speech	529,723	1,155,465	60	2026-09-05	apache-2.0	92
24	swiss-ai/Apertus-8B-Instruct-2509	swiss-ai	text-generation	508,878	3,673,347	491	2026-07-17	apache-2.0	92
25	LiquidAI/LFM2.5-1.2B-Instruct-GGUF	LiquidAI	text-generation	345,339	1,385,420	218	2026-08-24	other	92
26	nphSi/Z-Image-Lora	nphSi	text-to-image	265,540	1,396,374	149	2026-09-03	apache-2.0	92
27	Comfy-Org/Wan_2.2_ComfyUI_Repackaged	Comfy-Org	image-to-video	5,298,475	83,733,739	870	2026-08-17	apache-2.0	91
28	sentence-transformers/all-MiniLM-L12-v2	sentence-transformers	sentence-similarity	4,093,501	252,415,677	328	2026-03-31	apache-2.0	91
29	unsloth/gemma-4-12B-it-qat-GGUF	unsloth	any-to-any	1,336,575	2,278,607	494	2026-07-17	apache-2.0	91
30	prism-ml/Bonsai-27B-mlx-1bit	prism-ml	text-generation	1,238,560	2,529,744	237	2026-07-14	apache-2.0	91
31	prism-ml/Ternary-Bonsai-27B-mlx-2bit	prism-ml	text-generation	1,232,227	2,527,224	187	2026-07-14	apache-2.0	91
32	PaddlePaddle/PP-DocLayoutV3_safetensors	PaddlePaddle	object-detection	996,663	3,551,177	39	2026-07-08	apache-2.0	91
33	nvidia/parakeet-tdt-0.6b-v3	nvidia	automatic-speech-recognition	702,506	2,668,914	1,115	2026-08-05	cc-by-4.0	91
34	dots-studio/dots.mocr	dots-studio	image-text-to-text	627,770	2,313,846	169	2026-07-04	mit	91
35	cross-encoder/ms-marco-MiniLM-L6-v2	cross-encoder	text-ranking	87,478,105	543,283,080	316	2026-08-09	apache-2.0	90
36	Comfy-Org/stable-diffusion-v1-5-archive	Comfy-Org	other	17,379,879	41,553,675	128	2026-08-17	creativeml-openrail-m	90
37	Falconsai/nsfw_image_detection	Falconsai	image-classification	4,037,866	1,383,214,676	1,178	2026-09-07	apache-2.0	90
38	MahmoudAshraf/mms-300m-1130-forced-aligner	MahmoudAshraf	automatic-speech-recognition	2,777,141	75,853,998	103	2026-04-15	cc-by-nc-4.0	90
39	Comfy-Org/Wan_2.1_ComfyUI_repackaged	Comfy-Org	other	2,599,034	56,637,214	972	2026-08-17	apache-2.0	90
40	google/gemma-4-12B-it-qat-q4_0-gguf	google	any-to-any	780,934	1,819,649	290	2026-07-17	apache-2.0	90
41	unsloth/gemma-4-31B-it-qat-GGUF	unsloth	image-text-to-text	671,762	1,422,274	190	2026-07-17	apache-2.0	90
42	yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF	yuxinlu1	text-generation	646,541	1,525,823	1,553	2026-06-19	apache-2.0	90
43	zeroentropy/zerank-2-reranker	zeroentropy	text-ranking	644,922	1,567,093	117	2026-07-24	apache-2.0	90
44	palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4	palmfuture	image-text-to-text	405,300	1,642,801	29	2026-07-05	apache-2.0	90
45	DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF	DavidAU	image-text-to-text	396,985	1,656,423	430	2026-07-17	apache-2.0	90
46	mistralai/Ministral-3-14B-Instruct-2512	mistralai	other	396,159	2,220,841	320	2026-07-15	apache-2.0	90
47	protectai/unbiased-toxic-roberta-onnx	protectai	token-classification	271,155	1,582,815	7	2026-07-09	apache-2.0	90
48	nvidia/Qwen3.6-35B-A3B-NVFP4	nvidia	text-generation	10,030,544	32,868,681	592	2026-08-29	apache-2.0	89
49	Qwen/Qwen3.8-27B	Qwen	image-text-to-text	7,322,476	7,322,476	14,669	2026-08-14	apache-2.0	89
50	Qwen/Qwen3.5-35B-A3B-FP8	Qwen	image-text-to-text	1,457,050	10,359,960	156	2026-04-24	apache-2.0	89
51	microsoft/phi-4	microsoft	text-generation	711,395	14,228,376	2,299	2026-07-14	mit	89
52	deepseek-ai/DeepSeek-V4-Flash-DSpark	deepseek-ai	text-generation	697,663	1,435,950	271	2026-07-04	mit	89
53	unsloth/inkling-GGUF	unsloth	image-text-to-text	665,869	1,014,923	135	2026-07-16	apache-2.0	89
54	google/gemma-4-E4B-it-qat-q4_0-gguf	google	any-to-any	652,126	1,244,660	134	2026-07-17	apache-2.0	89
55	unsloth/gemma-4-E4B-it-qat-GGUF	unsloth	any-to-any	645,731	1,318,364	179	2026-07-17	apache-2.0	89
56	empero-ai/Qwythos-9B-v2-GGUF	empero-ai	image-text-to-text	643,526	1,241,421	267	2026-07-12	apache-2.0	89
57	biohub/ESMFold2-Experimental-Fast	biohub	other	629,347	1,140,558	0	2026-07-28	mit	89
58	biohub/ESMFold2-Experimental-Fast-Cutoff2025	biohub	other	588,154	1,072,759	0	2026-07-28	mit	89
59	AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4	AEON-7	image-text-to-text	584,762	1,317,652	72	2026-07-15	apache-2.0	89
60	google/gemma-4-E2B-it-qat-q4_0-gguf	google	any-to-any	545,379	1,170,313	124	2026-07-17	apache-2.0	89
61	google/gemma-4-E4B	google	any-to-any	544,839	3,414,085	416	2026-07-15	apache-2.0	89
62	Kijai/WanVideo_comfy_fp8_scaled	Kijai	other	543,118	8,426,406	734	2026-06-13	apache-2.0	89
63	DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF	DavidAU	image-text-to-text	482,407	2,047,448	738	2026-08-23	apache-2.0	89
64	biohub/ESMC-6B	biohub	fill-mask	2,588,347	7,104,085	31	2026-06-03	mit	88
65	nvidia/Gemma-4-26B-A4B-NVFP4	nvidia	text-generation	1,812,544	7,366,033	137	2026-05-11	apache-2.0	88
66	unsloth/Qwen3.6-35B-A3B-GGUF	unsloth	image-text-to-text	1,290,235	7,716,237	1,595	2026-04-20	apache-2.0	88
67	MiniMaxAI/MiniMax-M2.7	MiniMaxAI	text-generation	1,289,218	6,872,442	1,247	2026-04-20	other	88
68	unsloth/Qwen3.6-27B-GGUF	unsloth	image-text-to-text	1,221,320	5,572,541	955	2026-04-22	apache-2.0	88
69	k2-fsa/OmniVoice	k2-fsa	text-to-speech	1,194,599	8,136,312	1,369	2026-07-03	none	88
70	RunDiffusion/Juggernaut-XL-v9	RunDiffusion	text-to-image	851,323	7,543,768	435	2026-05-08	creativeml-openrail-m	88
71	DeepBeepMeep/Wan2.1	DeepBeepMeep	other	515,588	4,528,203	50	2026-09-04	none	88
72	unsloth/Kimi-K3-GGUF	unsloth	image-text-to-text	491,360	850,610	379	2026-08-07	other	88
73	Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF	Jackrong	image-text-to-text	397,849	841,197	335	2026-07-09	apache-2.0	88
74	biohub/ESMFold2-Fast	biohub	other	377,908	883,949	11	2026-07-28	mit	88
75	GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-GGUF	GnLOLot	text-generation	374,141	702,782	214	2026-07-13	apache-2.0	88
76	protectai/xlm-roberta-base-language-detection-onnx	protectai	text-classification	191,658	1,261,378	6	2026-07-09	mit	88
77	nomic-ai/nomic-embed-text-v1.5	nomic-ai	sentence-similarity	16,124,079	152,571,425	908	2026-04-07	apache-2.0	87
78	unsloth/Qwen3.8-27B-GGUF	unsloth	other	11,127,203	11,127,203	3,855	2026-08-20	apache-2.0	87
79	google/gemma-4-12B-it	google	any-to-any	3,076,461	10,489,657	1,546	2026-07-20	apache-2.0	87
80	Qwen/Qwen3-VL-Reranker-2B	Qwen	text-ranking	1,676,336	4,301,793	219	2026-04-16	apache-2.0	87
81	Qwen/Qwen3.5-122B-A10B-FP8	Qwen	image-text-to-text	1,503,077	7,379,506	115	2026-04-24	apache-2.0	87
82	Comfy-Org/flux2-dev	Comfy-Org	other	1,331,301	11,610,342	311	2026-08-17	other	87
83	litert-community/gemma-4-E2B-it-litert-lm	litert-community	other	1,100,398	5,169,078	422	2026-08-31	apache-2.0	87
84	sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP	sakamakismile	text-generation	962,607	3,328,036	81	2026-04-29	apache-2.0	87
85	HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive	HauhauCS	other	858,482	4,734,045	2,033	2026-06-05	apache-2.0	87
86	nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16	nvidia	text-generation	829,548	9,318,300	820	2026-08-24	other	87
87	unsloth/Qwen3.6-35B-A3B-MTP-GGUF	unsloth	image-text-to-text	808,947	3,265,663	911	2026-05-20	apache-2.0	87
88	bosonai/higgs-tts-2-3b-base	bosonai	text-to-speech	414,882	4,383,669	693	2026-06-25	other	87
89	Qwen/Qwen3.8-27B-FP8	Qwen	image-text-to-text	7,468,247	7,468,247	781	2026-08-14	apache-2.0	86
90	intfloat/multilingual-e5-base	intfloat	sentence-similarity	7,189,365	62,800,916	383	2026-04-02	mit	86
91	Comfy-Org/Krea-2	Comfy-Org	other	5,642,119	7,388,855	537	2026-08-17	other	86
92	MiniMaxAI/MiniMax-H3	MiniMaxAI	image-text-to-video	5,080,204	6,685,463	5,133	2026-08-13	other	86
93	datalab-to/chandra-ocr-2	datalab-to	image-text-to-text	2,917,927	10,928,997	499	2026-06-26	openrail	86
94	Qdrant/bm25	Qdrant	sentence-similarity	1,058,672	8,851,517	35	2026-08-20	apache-2.0	86
95	emrecan/bert-base-turkish-cased-mean-nli-stsb-tr	emrecan	sentence-similarity	438,773	7,957,822	52	2026-09-05	apache-2.0	86
96	sentence-transformers/all-MiniLM-L6-v2	sentence-transformers	sentence-similarity	254,324,263	3,822,982,827	5,773	2026-06-01	apache-2.0	85
97	google/gemma-4-E4B-it	google	any-to-any	4,770,720	30,405,025	1,542	2026-07-20	apache-2.0	85
98	iitolstykh/mivolo_v2	iitolstykh	other	2,647,955	26,340,270	32	2026-03-11	apache-2.0	85
99	ornith-ai/Ornith-1.0-35B	ornith-ai	text-generation	2,625,625	6,546,877	506	2026-06-25	mit	85
100	Comfy-Org/flux1-dev	Comfy-Org	other	511,225	5,668,616	669	2026-08-17	other	85
101	nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4	nvidia	text-generation	497,426	5,161,659	176	2026-08-24	other	85
102	Comfy-Org/HunyuanVideo_1.5_repackaged	Comfy-Org	other	494,913	6,386,875	97	2026-08-17	other	85
103	poolside/Laguna-S-2.1-NVFP4	poolside	text-generation	449,726	989,605	191	2026-08-25	openmdw-1.1	85
104	cyankiwi/Qwen3-VL-8B-Instruct-AWQ-4bit	cyankiwi	image-text-to-text	448,560	1,555,796	18	2026-07-21	apache-2.0	85
105	handy-computer/whisper-large-v3-turbo-gguf	handy-computer	automatic-speech-recognition	330,106	756,673	3	2026-07-21	apache-2.0	85
106	openbmb/MiniCPM-o-2_6	openbmb	any-to-any	275,363	5,178,795	1,299	2026-08-18	apache-2.0	85
107	trl-internal-testing/tiny-Qwen2ForCausalLM-2.5	trl-internal-testing	text-generation	18,620,166	111,590,553	24	2026-09-08	none	84
108	google/gemma-4-26B-A4B-it	google	image-text-to-text	8,867,625	57,598,600	1,486	2026-07-20	apache-2.0	84
109	ibm-granite/granite-embedding-small-english-r2	ibm-granite	feature-extraction	6,324,429	22,508,928	78	2026-01-21	apache-2.0	84
110	lmstudio-community/Qwen3.8-27B-MLX-4bit	lmstudio-community	image-text-to-text	4,275,484	4,275,484	40	2026-08-14	apache-2.0	84
111	lmstudio-community/Qwen3.8-27B-MLX-8bit	lmstudio-community	image-text-to-text	4,044,935	4,044,935	22	2026-08-14	apache-2.0	84
112	ornith-ai/Ornith-1.5-9B-GGUF	ornith-ai	text-generation	4,005,750	4,005,750	337	2026-08-24	mit	84
113	lmstudio-community/Qwen3.8-27B-MLX-6bit	lmstudio-community	image-text-to-text	3,978,884	3,978,884	12	2026-08-14	apache-2.0	84
114	lmstudio-community/Qwen3.8-27B-MLX-5bit	lmstudio-community	image-text-to-text	3,941,681	3,941,681	0	2026-08-14	apache-2.0	84
115	Qwen/Qwen3-TTS-12Hz-1.7B-Base	Qwen	other	3,613,870	17,305,054	509	2026-01-23	apache-2.0	84
116	ornith-ai/Ornith-1.5-35B-A3B-GGUF	ornith-ai	text-generation	3,539,705	3,539,705	392	2026-08-24	mit	84
117	unsloth/Qwen3.8-27B-NVFP4	unsloth	other	3,448,482	3,448,482	444	2026-09-08	apache-2.0	84
118	lmstudio-community/Qwen3.8-27B-GGUF	lmstudio-community	other	2,927,776	2,927,776	45	2026-08-14	apache-2.0	84
119	JonathanColetti/Qwen3.8-27B-Uncensored-GGUF	JonathanColetti	text-generation	2,812,743	2,812,743	1,051	2026-08-29	apache-2.0	84
120	baidu/Unlimited-OCR	baidu	image-text-to-text	2,724,383	7,404,462	4,212	2026-07-29	mit	84
121	lmstudio-community/gemma-4-E4B-it-MLX-4bit	lmstudio-community	any-to-any	1,173,873	5,155,322	24	2026-07-23	apache-2.0	84
122	lmstudio-community/gemma-4-E4B-it-MLX-8bit	lmstudio-community	any-to-any	1,148,910	4,990,587	10	2026-07-23	apache-2.0	84
123	lmstudio-community/gemma-4-E4B-it-MLX-5bit	lmstudio-community	any-to-any	1,141,444	4,621,542	0	2026-07-23	apache-2.0	84
124	lmstudio-community/gemma-4-E4B-it-MLX-6bit	lmstudio-community	any-to-any	1,139,401	4,926,154	3	2026-07-23	apache-2.0	84
125	ZhengPeng7/BiRefNet	ZhengPeng7	image-segmentation	1,009,624	19,353,123	635	2026-02-04	mit	84
126	unsloth/gemma-4-26B-A4B-it-qat-GGUF	unsloth	image-text-to-text	697,754	2,242,602	412	2026-07-17	apache-2.0	84
127	lmstudio-community/gemma-4-E4B-it-GGUF	lmstudio-community	other	621,035	4,972,517	69	2026-07-20	apache-2.0	84
128	Qwen/Qwen3.8-Flash-Next	Qwen	image-text-to-text	564,079	564,079	5,085	2026-08-27	other	84
129	unsloth/gemma-4-E4B-it-GGUF	unsloth	image-text-to-text	536,020	5,179,635	616	2026-07-17	apache-2.0	84
130	vidore/colqwen2.5-v0.2	vidore	visual-document-retrieval	328,936	2,814,216	101	2026-08-20	mit	84
131	google/tipsv2-so400m14	google	zero-shot-image-classification	272,258	597,001	19	2026-08-17	apache-2.0	84
132	Comfy-Org/flux1-schnell	Comfy-Org	other	243,571	4,027,959	282	2026-08-17	apache-2.0	84
133	OpenMOSS-Team/MOSS-Transcribe-Diarize	OpenMOSS-Team	audio-text-to-text	236,590	495,640	425	2026-09-02	apache-2.0	84
134	google/gemma-4-31B-it	google	image-text-to-text	8,744,810	56,221,699	3,757	2026-07-20	apache-2.0	83
135	Qwen/Qwen3.5-2B	Qwen	image-text-to-text	3,791,906	14,945,042	386	2026-03-02	apache-2.0	83
136	audio-cpp/audio.cpp-gguf	audio-cpp	text-to-speech	2,541,332	2,673,484	118	2026-09-09	other	83
137	cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF	cdiamond	image-text-to-text	2,480,386	2,480,386	8	2026-08-17	apache-2.0	83
138	huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF	huihui-ai	image-text-to-text	2,441,040	2,441,040	654	2026-09-08	apache-2.0	83
139	RadixArk/Qwen3.8-27B-NVFP4	RadixArk	image-text-to-text	2,293,133	2,293,133	89	2026-08-22	apache-2.0	83
140	mudler/Laguna-XS-2.1-APEX-GGUF	mudler	other	2,143,266	2,583,411	13	2026-08-17	openmdw-1.1	83
141	HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF	HauhauCS	image-text-to-text	1,908,917	1,908,917	1,080	2026-08-17	apache-2.0	83
142	nvidia/nemotron-3.5-asr-streaming-0.6b	nvidia	automatic-speech-recognition	958,097	2,363,755	1,100	2026-09-10	other	83
143	handy-computer/cohere-transcribe-03-2026-gguf	handy-computer	automatic-speech-recognition	948,079	2,372,468	3	2026-06-28	apache-2.0	83
144	Comfy-Org/vae-text-encorder-for-flux-klein-9b	Comfy-Org	other	897,529	2,008,803	198	2026-08-17	other	83
145	jinaai/jina-reranker-v3	jinaai	text-ranking	780,543	5,892,022	145	2026-08-10	cc-by-nc-4.0	83
146	cyankiwi/Qwen3.5-4B-AWQ-4bit	cyankiwi	image-text-to-text	723,019	3,866,369	21	2026-07-21	apache-2.0	83
147	vidore/colqwen2-v1.0	vidore	visual-document-retrieval	471,319	2,241,210	121	2026-08-20	apache-2.0	83
148	handy-computer/Voxtral-Mini-4B-Realtime-2602-gguf	handy-computer	automatic-speech-recognition	436,694	1,059,542	3	2026-06-28	apache-2.0	83
149	nlpai-lab/KURE-v1	nlpai-lab	feature-extraction	383,416	2,304,222	94	2026-09-07	mit	83
150	Comfy-Org/ace_step_1.5_ComfyUI_files	Comfy-Org	other	376,520	2,298,501	167	2026-08-17	apache-2.0	83
151	Comfy-Org/HunyuanVideo_repackaged	Comfy-Org	other	349,560	2,148,434	244	2026-08-17	other	83
152	nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4	nvidia	text-generation	303,089	1,231,447	327	2026-08-24	other	83
153	giacomoarienti/nsfw-classifier	giacomoarienti	image-classification	255,679	2,247,869	56	2026-09-03	cc-by-nc-nd-4.0	83
154	ornith-ai/Ornith-1.0-9B-GGUF	ornith-ai	text-generation	3,989,012	10,401,936	665	2026-06-25	mit	82
155	handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf	handy-computer	automatic-speech-recognition	1,813,301	4,708,779	9	2026-06-29	other	82
156	0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF	0bserverx	text-generation	1,721,344	1,721,344	437	2026-08-20	apache-2.0	82
157	nvidia/Cosmos3-Edge	nvidia	other	1,652,551	1,781,023	198	2026-08-26	other	82
158	unsloth/Qwen3.5-9B-GGUF	unsloth	image-text-to-text	1,634,867	8,112,272	898	2026-03-02	apache-2.0	82
159	handy-computer/parakeet-unified-en-0.6b-gguf	handy-computer	automatic-speech-recognition	1,549,452	4,170,009	6	2026-09-10	cc-by-4.0	82
160	cyankiwi/Qwen3.8-27B-AWQ-INT4	cyankiwi	image-text-to-text	1,492,190	1,492,190	94	2026-08-15	apache-2.0	82
161	lightx2v/Minimax-h3-Turbo	lightx2v	image-to-video	1,372,782	1,457,062	890	2026-09-10	apache-2.0	82
162	ggml-org/Qwen3.8-27B-GGUF	ggml-org	image-text-to-text	1,372,097	1,372,097	78	2026-08-14	apache-2.0	82
163	mudler/KAT-Coder-V2.5-Dev-APEX-GGUF	mudler	other	1,326,177	1,632,388	55	2026-08-17	apache-2.0	82
164	nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4	nvidia	text-generation	1,321,195	1,361,309	410	2026-09-10	other	82
165	Inferact/Qwen3.8-27B-NVFP4	Inferact	image-text-to-text	1,218,591	1,218,591	17	2026-08-14	apache-2.0	82
166	OBLITERATUS/Qwen3.8-27B-OBLITERATED	OBLITERATUS	text-generation	1,133,995	1,133,995	1,158	2026-08-24	apache-2.0	82
167	LiquidAI/LFM2.5-2.6B-GGUF	LiquidAI	text-generation	1,098,072	1,306,025	336	2026-08-24	other	82
168	LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-GGUF	LuffyTheFox	image-text-to-text	843,376	1,454,882	608	2026-09-07	apache-2.0	82
169	Comfy-Org/ltx-2	Comfy-Org	other	783,095	1,755,954	148	2026-08-17	other	82
170	google/gemma-4-31B	google	image-text-to-text	687,341	3,478,355	525	2026-07-15	apache-2.0	82
171	Serveurperso/Qwen3-TTS-GGUF	Serveurperso	text-to-speech	672,661	1,143,655	43	2026-09-09	apache-2.0	82
172	wangzhang/gemma-4-31B-it-abliterated	wangzhang	other	546,247	1,743,387	32	2026-08-29	apache-2.0	82
173	handy-computer/whisper-medium-gguf	handy-computer	automatic-speech-recognition	515,797	1,255,059	0	2026-06-28	apache-2.0	82
174	Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF	Jackrong	image-text-to-text	485,982	1,323,621	238	2026-07-09	apache-2.0	82
175	Qwen/Qwen3.5-9B-Base	Qwen	image-text-to-text	469,875	1,687,527	105	2026-04-23	apache-2.0	82
176	Lightricks/LTX-2	Lightricks	image-to-video	361,535	8,478,496	1,780	2026-08-04	other	82
177	mistralai/Devstral-Small-2-24B-Instruct-2512	mistralai	other	247,647	2,613,806	659	2026-07-15	apache-2.0	82
178	openbmb/MiniCPM-V-4_5	openbmb	image-text-to-text	232,651	1,829,723	1,099	2026-08-18	apache-2.0	82
179	Comfy-Org/z_image	Comfy-Org	other	194,063	1,467,337	269	2026-08-17	apache-2.0	82
180	bigscience/bloom	bigscience	text-generation	0	4,883,402	5,043	2026-07-29	bigscience-bloom-rail-1.0	82
181	openbmb/MiniCPM-V-4	openbmb	image-text-to-text	0	1,261,885	464	2026-08-18	apache-2.0	82
182	nvidia/canary-1b-v2	nvidia	automatic-speech-recognition	0	1,697,856	419	2026-08-31	cc-by-4.0	82
183	Comfy-Org/stable-diffusion-3.5-fp8	Comfy-Org	other	0	1,409,193	238	2026-08-17	other	82
184	Qwen/Qwen3-Embedding-0.6B	Qwen	feature-extraction	7,806,497	84,274,770	1,190	2026-04-20	apache-2.0	81
185	Qwen/Qwen3.5-4B	Qwen	image-text-to-text	7,205,448	43,611,867	906	2026-03-02	apache-2.0	81
186	Qwen/Qwen3.6-27B-FP8	Qwen	image-text-to-text	6,953,349	33,241,219	355	2026-04-24	apache-2.0	81
187	google/gemma-4-E2B-it	google	any-to-any	3,188,731	17,527,634	950	2026-07-20	apache-2.0	81
188	Qwen/Qwen3.5-35B-A3B	Qwen	image-text-to-text	2,211,686	18,721,824	1,501	2026-04-24	apache-2.0	81
189	unsloth/Qwen3.8-Flash-Next-GGUF	unsloth	image-text-to-text	1,053,823	1,053,823	880	2026-09-02	other	81
190	zai-org/GLM-5.3-Flash	zai-org	image-text-to-text	1,023,103	1,023,103	2,232	2026-09-07	mit	81
191	ornith-ai/Ornith-1.5-397B-GGUF	ornith-ai	text-generation	1,019,173	1,019,173	35	2026-08-24	mit	81
192	Bahushruth/Qwen3.6-35B-A3B-abliterated-v4	Bahushruth	text-generation	935,126	2,137,732	8	2026-07-03	apache-2.0	81
193	ReliquaryForge/qwen3-4b-base-dapo-v4	ReliquaryForge	text-generation	889,036	889,036	2	2026-09-11	apache-2.0	81
194	ornith-ai/Ornith-1.5-35B-A3B-NVFP4	ornith-ai	text-generation	824,957	824,957	47	2026-08-26	mit	81
195	Comfy-Org/MiniMax-Music-3	Comfy-Org	other	814,913	814,913	228	2026-08-17	apache-2.0	81
196	openbmb/MiniCPM-o-4_5	openbmb	any-to-any	694,166	2,651,869	1,480	2026-08-18	apache-2.0	81
197	unsloth/MiniMax-H3-GGUF	unsloth	image-text-to-video	689,374	790,614	272	2026-08-14	other	81
198	handy-computer/parakeet-tdt-0.6b-v3-gguf	handy-computer	automatic-speech-recognition	591,930	1,465,775	6	2026-06-28	cc-by-4.0	81
199	unsloth/gemma-4-31B-it-GGUF	unsloth	image-text-to-text	454,123	4,817,430	602	2026-07-17	apache-2.0	81
200	LiquidAI/LFM2.5-8B-A1B-GGUF	LiquidAI	text-generation	453,013	824,734	296	2026-08-24	other	81
201	rdtand/Qwen3.6-27B-PrismaSCOUT-Blackwell-NVFP4-BF16-vllm	rdtand	other	417,330	814,003	43	2026-08-24	apache-2.0	81
202	mudler/gemma-4-26B-A4B-it-APEX-GGUF	mudler	other	405,966	923,365	82	2026-08-17	gemma	81
203	mudler/Qwen3.6-35B-A3B-APEX-GGUF	mudler	other	386,864	863,558	198	2026-08-17	apache-2.0	81
204	athrael-soju/colqwen3.5-4.5B-v3	athrael-soju	visual-document-retrieval	382,467	941,074	17	2026-08-21	apache-2.0	81
205	lmstudio-community/gemma-4-E2B-it-MLX-4bit	lmstudio-community	any-to-any	209,154	883,782	1	2026-07-23	apache-2.0	81
206	lmstudio-community/gemma-4-E2B-it-MLX-8bit	lmstudio-community	any-to-any	200,928	841,220	1	2026-07-23	apache-2.0	81
207	lmstudio-community/gemma-4-E2B-it-MLX-6bit	lmstudio-community	any-to-any	200,362	832,636	0	2026-07-23	apache-2.0	81
208	lmstudio-community/gemma-4-E2B-it-MLX-5bit	lmstudio-community	any-to-any	200,132	748,170	0	2026-07-23	apache-2.0	81
209	ggml-org/gpt-oss-120b-GGUF	ggml-org	text-generation	196,529	2,634,351	76	2026-07-28	apache-2.0	81
210	microsoft/Fara-7B	microsoft	image-text-to-text	0	812,661	620	2026-08-12	mit	81
211	LiquidAI/LFM2-1.2B	LiquidAI	text-generation	0	3,087,131	366	2026-08-05	other	81
212	intfloat/multilingual-e5-small	intfloat	sentence-similarity	12,369,093	109,949,447	401	2026-04-02	mit	80
213	autogluon/chronos-2	autogluon	time-series-forecasting	9,038,338	69,716,219	49	2026-06-05	apache-2.0	80
214	cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit	cyankiwi	image-text-to-text	2,256,988	19,517,802	97	2026-09-10	apache-2.0	80
215	unsloth/Inkling-Small-GGUF	unsloth	image-text-to-text	1,413,473	1,670,549	84	2026-07-31	apache-2.0	80
216	mudler/ced-gguf	mudler	audio-classification	1,408,776	1,784,687	3	2026-06-21	apache-2.0	80
217	farbodtavakkoli/OTel-LLM-27B-IT	farbodtavakkoli	text-generation	1,356,059	2,029,181	0	2026-06-23	apache-2.0	80
218	sahilchachra/Unlimited-OCR-AWQ	sahilchachra	image-text-to-text	1,335,974	1,915,542	2	2026-06-23	mit	80
219	Qwen/Qwen3-VL-Embedding-2B	Qwen	sentence-similarity	1,307,966	10,480,242	449	2026-04-16	apache-2.0	80
220	raxcore-dev/Rax-4.5	raxcore-dev	image-text-to-text	1,059,991	1,458,605	5	2026-07-13	apache-2.0	80
221	ggml-org/gemma-4-E4B-it-GGUF	ggml-org	any-to-any	994,313	1,421,234	85	2026-07-26	apache-2.0	80
222	Abiray/MiniMax-H3-GGUF	Abiray	image-to-video	952,223	1,575,071	132	2026-08-08	other	80
223	gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090	gittensor-model-hub	image-text-to-text	714,020	714,020	154	2026-09-10	apache-2.0	80
224	empero-ai/Qwen3.8-4B-Distill-GGUF	empero-ai	text-generation	640,502	640,502	109	2026-08-16	apache-2.0	80
225	empero-ai/Qwen3.8-2B-Distill-GGUF	empero-ai	text-generation	631,633	631,633	130	2026-08-16	apache-2.0	80
226	ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF	ISTA-DASLab	image-text-to-text	614,850	614,850	806	2026-09-02	apache-2.0	80
227	empero-ai/Qwen3.8-9B-Distill-GGUF	empero-ai	text-generation	580,426	580,426	227	2026-08-16	apache-2.0	80
228	meta-models/Muse-Glimmer-30B	meta-models	image-text-to-text	578,047	693,279	1,877	2026-08-11	apache-2.0	80
229	RedHatAI/gemma-4-12B-it-FP8-Dynamic	RedHatAI	any-to-any	573,909	626,843	8	2026-08-13	apache-2.0	80
230	DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF	DavidAU	image-text-to-text	553,017	553,017	310	2026-08-23	apache-2.0	80
231	zai-org/GLM-5.3	zai-org	text-generation	552,019	552,019	1,800	2026-09-04	other	80
232	AnkitAI/Parable-Qwen3-8B-Claude-Fable-5-GGUF	AnkitAI	text-generation	551,982	716,121	3	2026-08-20	apache-2.0	80
233	ornith-ai/Ornith-1.5-397B	ornith-ai	text-generation	547,857	547,857	85	2026-08-23	mit	80
234	z-lab/Qwen3.8-27B-DFlash2-GGUF	z-lab	text-generation	541,874	541,874	123	2026-08-24	apache-2.0	80
235	LiquidAI/LFM2.5-230M-GGUF	LiquidAI	text-generation	522,161	576,778	98	2026-08-24	other	80
236	empero-ai/Qwen3.8-27B-Ridge-GGUF	empero-ai	image-text-to-text	518,360	518,360	339	2026-08-15	apache-2.0	80
237	DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF	DavidAU	image-text-to-text	517,644	517,644	453	2026-09-08	apache-2.0	80
238	nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16	nvidia	text-generation	503,676	525,874	204	2026-08-24	other	80
239	Comfy-Org/Wan-Animate-2	Comfy-Org	other	496,398	610,170	48	2026-08-17	apache-2.0	80
240	ornith-ai/Ornith-1.5-9B-NVFP4	ornith-ai	text-generation	494,394	494,394	10	2026-08-26	mit	80
241	google/timesfm-3.0-pytorch	google	time-series-forecasting	483,787	483,787	718	2026-09-02	other	80
242	Comfy-Org/gemma-4	Comfy-Org	other	458,543	531,158	74	2026-08-17	apache-2.0	80
243	cointegrated/rubert-tiny-toxicity	cointegrated	text-classification	439,907	1,771,314	48	2026-08-03	mit	80
244	Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF	Jackrong	image-text-to-text	418,132	1,515,691	314	2026-07-04	apache-2.0	80
245	meta-models/Muse-Glimmer-30B-GGUF	meta-models	image-text-to-text	389,466	517,730	333	2026-08-18	apache-2.0	80
246	AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF	AnkitAI	text-generation	387,351	539,951	1	2026-08-21	apache-2.0	80
247	Comfy-Org/Qwen3-VL	Comfy-Org	other	353,687	590,175	59	2026-08-17	apache-2.0	80
248	GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF	GnLOLot	text-generation	344,932	766,830	329	2026-07-13	apache-2.0	80
249	Comfy-Org/SeedVR2	Comfy-Org	other	336,824	509,643	84	2026-08-17	apache-2.0	80
250	Comfy-Org/SCAIL-2	Comfy-Org	other	332,617	599,666	112	2026-08-17	mit	80
251	Comfy-Org/vae-text-encorder-for-flux-klein-4b	Comfy-Org	other	329,072	501,968	90	2026-08-17	apache-2.0	80
252	mudler/Qwen3.5-35B-A3B-APEX-GGUF	mudler	text-generation	280,263	600,868	94	2026-08-17	apache-2.0	80
253	Comfy-Org/ltx-2.3	Comfy-Org	other	237,266	548,243	40	2026-08-17	other	80
254	google/tipsv2-b14	google	zero-shot-image-classification	229,792	496,501	124	2026-08-17	apache-2.0	80
255	Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled	Jackrong	image-text-to-text	0	1,257,779	2,944	2026-07-07	apache-2.0	80
256	zai-org/GLM-5	zai-org	text-generation	0	1,340,133	2,119	2026-08-11	mit	80
257	dealignai/Gemma-4-31B-JANG_4M-CRACK	dealignai	image-text-to-text	0	364,603	1,711	2026-09-08	gemma	80
258	froggeric/Qwen-Fixed-Chat-Templates	froggeric	other	0	0	1,646	2026-09-04	apache-2.0	80
259	deepseek-ai/DeepSeek-V4.1-Flash	deepseek-ai	image-text-to-text	0	6	1,483	2026-09-10	mit	80
260	vidore/colpali	vidore	visual-document-retrieval	0	463,650	487	2026-08-20	mit	80
261	google/gemma-4-E2B	google	any-to-any	0	2,026,327	468	2026-07-15	apache-2.0	80
262	naver-hyperclovax/HyperCLOVAX-SEED-Think-32B	naver-hyperclovax	text-generation	0	633,805	405	2026-09-10	other	80
263	Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-GGUF	Jackrong	image-text-to-text	0	1,395,207	353	2026-07-09	apache-2.0	80
264	nvidia/canary-1b-flash	nvidia	automatic-speech-recognition	0	2,081,526	279	2026-06-29	cc-by-4.0	80
265	Qwen/Qwen3-ASR-1.7B	Qwen	automatic-speech-recognition	2,902,342	15,140,185	1,085	2026-01-30	apache-2.0	79
266	RadixArk/Kimi-K3-DSpark	RadixArk	text-generation	2,645,687	4,734,235	55	2026-08-16	none	79
267	Qwen/Qwen3-Reranker-4B	Qwen	text-ranking	2,546,793	12,078,065	156	2026-04-16	apache-2.0	79
268	trl-internal-testing/tiny-Qwen3ForCausalLM	trl-internal-testing	text-generation	1,902,827	5,758,799	1	2026-09-08	none	79
269	unsloth/Qwen3.6-35B-A3B-NVFP4	unsloth	image-text-to-text	1,640,310	4,671,845	117	2026-07-12	apache-2.0	79
270	Kijai/WanVideo_comfy	Kijai	other	1,458,040	71,190,118	2,505	2026-06-13	none	79
271	Lightricks/LTX-2.3	Lightricks	image-to-video	1,194,236	11,654,832	1,881	2026-08-27	other	79
272	trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration	trl-internal-testing	image-text-to-text	1,012,968	6,116,756	0	2026-07-22	none	79
273	LocalAI-io/privacy-filter-nemotron-GGUF	LocalAI-io	token-classification	958,178	1,211,894	0	2026-06-19	apache-2.0	79
274	bartowski/endless-frontier_BigBang-v1-GGUF	bartowski	image-text-to-text	951,055	1,005,436	30	2026-08-07	apache-2.0	79
275	trl-internal-testing/tiny-GptOssForCausalLM	trl-internal-testing	text-generation	941,095	5,245,486	4	2026-09-08	none	79
276	Qwen/Qwen3.5-397B-A17B-FP8	Qwen	image-text-to-text	790,614	6,200,920	183	2026-04-24	apache-2.0	79
277	FINAL-Bench/POCKET-35B-GGUF	FINAL-Bench	text-generation	785,806	977,963	72	2026-07-29	apache-2.0	79
278	unsloth/Muse-Glimmer-30B-GGUF	unsloth	image-text-to-text	771,700	1,101,365	524	2026-08-10	apache-2.0	79
279	protoLabsAI/Ornith-1.0-35B-FP8	protoLabsAI	text-generation	708,909	1,007,208	8	2026-07-03	mit	79
280	prism-ml/Ternary-Bonsai-27B-gguf	prism-ml	text-generation	645,610	1,550,067	1,308	2026-08-31	apache-2.0	79
281	unsloth/gemma-4-E2B-it-qat-GGUF	unsloth	any-to-any	590,904	969,077	73	2026-07-17	apache-2.0	79
282	google/gemma-4-26B-A4B-it-qat-q4_0-gguf	google	image-text-to-text	570,315	1,142,260	164	2026-07-17	apache-2.0	79
283	google/gemma-4-31B-it-qat-q4_0-gguf	google	image-text-to-text	480,399	948,408	127	2026-07-17	apache-2.0	79
284	bartowski/Qwen3.8-27B-GGUF	bartowski	image-text-to-text	448,655	448,655	135	2026-08-14	apache-2.0	79
285	unsloth/Qwen3.8-27B	unsloth	other	444,992	444,992	47	2026-08-14	apache-2.0	79
286	philbert440/Qwen3.8-27B-W4A16-AWQ	philbert440	image-text-to-text	436,336	436,336	25	2026-08-15	apache-2.0	79
287	trl-internal-testing/tiny-LlamaForCausalLM-3.2	trl-internal-testing	text-generation	435,064	6,470,003	1	2026-09-10	none	79
288	google/gemma-4-E2B-it-qat-w4a16-ct	google	any-to-any	425,544	1,436,388	10	2026-07-20	apache-2.0	79
289	deepgrove/maple-preview-GGUF	deepgrove	text-generation	423,694	450,051	63	2026-08-15	mit	79
290	deepseek-ai/DeepSeek-V4-Flash-Vision-Exp	deepseek-ai	image-text-to-text	400,892	400,892	860	2026-09-01	mit	79
291	webAI-Official/TwIL-LM3	webAI-Official	text-generation	391,785	396,034	87	2026-09-09	other	79
292	unsloth/Kimi-K2.7-Code-GGUF	unsloth	image-text-to-text	374,402	986,280	192	2026-06-19	other	79
293	ornith-ai/Ornith-1.5-9B	ornith-ai	text-generation	373,653	373,653	291	2026-08-23	mit	79
294	RadixArk/Qwen3.8-27B-DSpark	RadixArk	text-generation	368,359	368,359	78	2026-08-29	other	79
295	incoai/Qwen3.8-27B-DFlash2	incoai	text-generation	339,853	339,853	230	2026-08-19	apache-2.0	79
296	Blackfrost-AI/Qwen3.8-27B-ABLITERATED-GGUF	Blackfrost-AI	image-text-to-text	336,269	336,269	230	2026-08-22	apache-2.0	79
297	Qwen/Qwen3.8-Flash-Next-FP8	Qwen	image-text-to-text	334,466	334,466	197	2026-08-31	other	79
298	ornith-ai/Ornith-1.5-397B-NVFP4	ornith-ai	text-generation	329,530	329,530	18	2026-08-29	mit	79
299	tencent/Hy-MT2-1.8B-GGUF	tencent	other	328,237	416,635	183	2026-09-08	apache-2.0	79
300	nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark	nvidia	text-generation	326,043	327,841	28	2026-09-01	other	79

How to Use and Fine-Tune a Hugging Face Model

Every model on this list can be loaded directly with the transformers library, using AutoModel.from_pretrained(“org/model-id”), or downloaded standalone with huggingface_hub’s snapshot_download. Click a model’s name to open its official Hugging Face page, where the model card documents exact usage, required libraries, and any license terms you need to accept first.

Fine-tuning is the other half of the question. You rarely need it: for most tasks a retrieval pipeline over your own documents beats a fine-tune, costs less, and updates the moment your documents change.

When you genuinely do need the model to learn a behaviour rather than a fact, the standard route is a parameter-efficient method such as LoRA, which trains a small adapter on top of the frozen pre-trained weights instead of retraining the whole model, so it fits on hardware you already have and produces a file measured in megabytes rather than gigabytes. Start from the smallest model that could plausibly work, because a fine-tuned 4B that runs cheaply usually beats a 30B you cannot afford to serve.

A short checklist before you commit to one in production:

1. Check the licence on the model card, not the tag. 9 models here state none at all, and “other” covers custom terms.
2. Prefer downloads over likes for infrastructure choices. The most-liked model on this list is #18 by downloads.
3. Check the last-modified date, remembering it moves on README edits too.
4. Match the format to your hardware. 47% of this list exists because the original weights did not fit somewhere; a GGUF or AWQ build is often what you want.
5. Watch for gated models if you are automating downloads, since they need terms accepted first.
6. Test on your own data before you standardise. Every ranking on this page, including mine, is a prior rather than an answer.

Where Hugging Face datasets fit alongside the models

A model on the Hub is the trained artefact. A dataset is the material it learned from or is measured against: a published collection of text, images, audio or tabular rows, with a dataset card documenting its schema, splits and licence.

Pick a model from the list above and you will usually end up in the dataset half of the Hub as well, either to fine-tune it or to score it against a benchmark. Anything there loads in a single call with the `datasets` library, `load_dataset(“org/dataset-id”)`, which handles the download, caching and format conversion for you. Click through to the dataset card first, because that is where the schema, the splits and any terms you have to accept are documented.

We ran this same ranking method over Hub datasets: the same adoption floor of 1,000 downloads in 30 days or 50 likes, and the same quality score. It returned 150 datasets, 26,092,772 downloads in 30 days and 174,550,323 all-time, from 132 publishers.

Everything in this section comes from that run, pulled 4 August 2026. Unlike the model tables above it is a fixed snapshot rather than a figure that refreshes, so read it as a point-in-time reading of the data layer.

The 10 most-downloaded Hugging Face datasets

#	Dataset	Publisher	Downloads (30d)	Licence
1	fineweb-tokenized	anisoleai	4,557,390	ODC-BY
2	video-vec2wav2-tokenizer	k9cli	2,058,719	None stated
3	hd_tmp	ayuo	1,472,506	None stated
4	PhysicalAI-Robotics-GR00T-X-Embodiment-Sim	nvidia	1,292,603	CC-BY-4.0
5	ubuntu_osworld_file_cache	xlangai	1,203,942	Apache-2.0
6	gsm8k	openai	936,722	MIT
7	LLaVA-OneVision-1.5-Mid-Training-85M	mvp-lab	698,674	Apache-2.0
8	KakologArchives	KakologArchives	696,977	MIT
9	results	mteb	535,391	None stated
10	figofigofigofigo	Dagonulca	532,006	None stated

Those ten take 53.6% of all downloads across the 150, close to the 53.9% the top ten models take above. Adoption on the Hub is top-heavy on both sides of it, and for the same reason: a handful of assets get written into tutorials, training recipes and CI pipelines, and then nobody swaps them out.

Dataset licensing is far messier than model licensing

This is the finding worth carrying over from the dataset side, because it changes what you are allowed to ship. 75% of the 300 models above are Apache-2.0 or MIT. Only 32% of the 150 datasets are:

• 48 of 150 (32%) Apache-2.0 or MIT.
• 42 under CC0 or an open Creative Commons licence, which permits commercial use with attribution.
• 17 explicitly non-commercial (cc-by-nc-4.0 or cc-by-nc-sa-4.0). Training a model you intend to sell on one of these is the precise use they forbid, and it is a much larger group proportionally than the four non-commercial models above.
• 9 state no licence at all, plus 2 reporting “unknown”. No stated licence means no permission granted.
• 7 are gated, so an automated download fails until somebody accepts the terms in a browser.

Maintenance looks similar to the model side: a median of 40 days since the last update, 67 of 150 touched within 30 days, and not one left untouched for over a year. Quality scores land in a 74 to 90 band with a median of 81, slightly below the model band because dataset repos collect fewer likes for the same amount of real use.

Robotics and simulation is the fastest-moving dataset category

Ten robotics and simulation datasets averaged 203,048 downloads each over 30 days, against 97,668 for the 29 vision and video datasets and less again for everything else. The fastest growers are almost all embodied or agentic benchmarks: RoboDojo at +190% over 30 days, BEHAVIOR-1K’s 2026 challenge demos at +176%, Nebius’s SWE-rebench at +156%. NVIDIA’s GR00T embodiment simulation set is the only robotics entry in the download top five.

The same shape shows up on the model side of this report. The smallest task groups by model count are not the smallest by downloads, and the categories that look marginal on a count of repositories are often the ones being pulled hardest.

What this data says about the Hugging Face Hub

Five conclusions, all checkable against the tables above:

1. The Hub runs on embeddings. 18 Embeddings & Retrieval models take 47.6% of downloads, roughly 15x the downloads per model of the most crowded category.
2. Downloads are a defaults game. One compact model from years ago holds 29.8% of all downloads, because it is what the tutorials use.
3. Attention and use are different axes. The most-liked model ranks #18 by downloads.
4. The open-weights market has split in two. 28 chat models under 10B for the hardware people own, 78 in the 20B to 40B band that mixture-of-experts made practical, and one model in between.
5. A large slice of the ecosystem is repackaging, not new models. 47% of the list is quantized or converted re-uploads, and 9 models with no licence stated are the group that can actually cause you a problem.

“The number that reframed this for me is that one small embedding model from a few years ago pulls more downloads than every image generator, speech model and classifier on the list combined. Everyone argues about which chat model is best. Meanwhile the thing quietly running in production is a tiny sentence-transformer that got written into a tutorial once and never got replaced. If you want to know what an ecosystem actually depends on, count what it downloads, not what it upvotes.”
Alston Antony, founder of zplatform.ai and Senior Digital Marketing Manager at Brainstorm Force

Building AI into a product rather than picking a model? Our MCP servers directory ranks the connectors that let assistants act on real systems.

For AI inside your site, see the best AI WordPress plugins, which carries a full CVE audit. For AI in the browser, the AI Chrome extensions and AI Firefox add-ons reports use the same data-first approach. If a term on this page is unfamiliar, the AI glossary defines it, and more roundups sit in best AI tools.

Questions people ask about Hugging Face models

What are the best Hugging Face models?

By composite quality score the leader is argmaxinc/whisperkit-coreml at 100/100, and by raw downloads it is sentence-transformers/all-MiniLM-L6-v2 with 254,324,263 in 30 days. Which matters depends on the job: downloads indicate a proven, safe default, while the quality score weighs maintenance, growth and licensing alongside adoption. For a single pick per task, use the picks table at the top of this page.

Which Hugging Face models are the most downloaded?

sentence-transformers/all-MiniLM-L6-v2, a sentence-similarity model, with 254,324,263 downloads in the last 30 days, which is 29.8% of all downloads across this list. Then cross-encoder/ms-marco-MiniLM-L6-v2 at 87,478,105 and Comfy-Org/MiniMax-H3 at 18,813,055. The top three are 42.2% of the whole list between them, and two of the three are retrieval models rather than chat models, because embedding models get pulled into every build of every search and RAG pipeline that uses them.

What are the top open-source LLMs on Hugging Face right now?

Judged by downloads on this list, the open-weight chat models people actually pull sit overwhelmingly in the 20B to 40B band, led by Qwen/Qwen3.6-35B-A3B-FP8. Nearly all of the current leaders are mixture-of-experts builds, and a large share of their downloads go to quantized GGUF or FP4 re-uploads rather than the original weights, because that is the form that fits on consumer hardware.

How do I choose the best Hugging Face model for my use case?

Start from the task, not the leaderboard. Narrow to the Hub task tag that matches the job, filter to models whose licence permits what you are building, check the size against the memory you have, and then compare the shortlist on your own data.

The picks table above does the first three steps for nine common jobs. The fourth step is the one nobody can do for you, and it is the one that decides the outcome.

Are Hugging Face models free to use?

Most are. 75% of this list is Apache-2.0 or MIT, which permits commercial use with minimal conditions. But 9 models state no licence, which means no permission has been granted, and four are explicitly non-commercial. Always read the model card before shipping.

How many models are on the Hugging Face Hub?

Well over a million, the overwhelming majority of which are never downloaded by anyone but their author. This report ranks the 300 that clear an adoption floor of at least 1,000 downloads in 30 days or 50 likes, as of 11 September 2026.

How is this different from Hugging Face’s own trending page?

Hugging Face’s trending view surfaces short-term spikes. This list weighs sustained adoption, how actively a model is maintained, and whether it is properly licensed, so it favours models genuinely in use over models briefly in the news. Hugging Face’s internal trending score is deliberately excluded from the ranking because its scale is not documented.

Are Hugging Face datasets free to use?

Most are free to download, but the licensing is looser than on the model side. In our August 2026 run over Hub datasets, only 32% of the 150 ranked datasets were Apache-2.0 or MIT, 17 were explicitly non-commercial, 9 stated no licence at all, and 7 were gated behind terms somebody has to accept in a browser before the download works. Read the dataset card before you train anything you intend to sell.

How often is this list updated?

The dataset is rebuilt from the Hugging Face Hub API and republished here, with the pull date shown at the top. Growth figures compare each model against a stored snapshot from roughly 30 days earlier, so they depend on that history being kept rather than on any single refresh.

Sources, method and how to cite this data

Source: the free, public Hugging Face Hub API, which supplies downloads (30-day and all-time), likes, licence, author, gated status, task tag and last-modified date for every model, plus 6 retained snapshots used to compute the 30-day download trend.

Sample: 300 models, 853,841,158 downloads in 30 days, 8,382,324,609 all-time, 179,513 likes, 111 publishers, 8 task groups, 13 distinct licences.

Inclusion: automatic, with an adoption floor of at least 1,000 downloads in the last 30 days or at least 50 likes. The Hub hosts well over a million models and most have never been downloaded by anyone but their author, so ranking all of them would be meaningless. No hand-picking, and no model pays for placement.

Ranking formula: Quality Score = 40% adoption (downloads and likes, log-scaled) + 25% maintenance (how recently updated) + 20% growth (30-day downloads trend) + 15% trust (a stated licence, an identified author, and not being access-gated).

Picks by job: for each task the shortlist is the five most-downloaded models the Hub tags for it, and the pick is whichever of those five carries the highest quality score.

Parameter counts are read from the model id, which is the only place the Hub API exposes one. The weight-band table covers the 155 models the Hub tags as generative, 116 of which state a size, and an id naming both a total and an active-expert count is banded on the total.

Stated limitations: because of the adoption floor, quality scores sit in a narrow 79 to 100 band and should be read as an ordering within an already-filtered set. A model released too recently to clear the floor does not appear at all, however good it is.

Growth is null for 150 models too new to have 30-day history, and those are scored neutrally rather than penalised. Task groups follow the Hub’s own task tag, which files some chat models under vision because they also accept images.

The last-modified date moves on any repository change, documentation included, so it indicates attention rather than retraining. Hugging Face’s own trending score is reported by the API but excluded because its scale is undocumented.

Cite as: zplatform.ai, “Best Hugging Face Models: Ranked on Real Download Data,” data pulled 11 September 2026.

Every number here traces to a row in the tables above, which is the point of publishing the formula next to the ranking. The best Hugging Face model for you depends on the task, the hardware and the licence you can live with, and you should be able to check my working rather than take my word for it.