Report · Data from Feb 12 to Oct 1, 2026

What 231 days of Hugging Face downloads say

Research by Azumo · Published 2 October 2026

We recorded the most-downloaded models on Hugging Face, the public hub for open AI models, every day from 12 February to 1 October 2026.

Read the summary

LLMs and embedding models went from about 45% to about 58% of downloads, and converted copies of LLMs and two embedding and reranking repos carry most of that gain. Vision and text classification fell from about 25% to about 16%, losing almost two-fifths of their share. For embedding, vision and text classification models, three in four downloads still go to models from 2023 or earlier.

Explore the findings

28%

of top-5,000 downloads go to embedding and reranker models, close to LLMs' 30%, from about 8% of the slots

3 in 4

embedding, vision and text classification downloads go to models from 2023 or earlier; for LLMs it is 7%

25% to 16%

vision and text classification's share of downloads since mid-March

About half

of image and video model downloads, as Hugging Face counts them, go to ComfyUI repackages of the labs' models

Download counts include automated activity and repeat requests. We compare shares and state the time window for each finding.

How to read the data

The market by model type

Model types follow our trends pages: each repo is placed by its Hugging Face task tag. About 11% of downloads go to repos with no task tag, such as ComfyUI repackages, Google's ELECTRA encoder and YOLO detectors, and those sit in no type.

Share by model type since March

Vision and text classification models were about 25% of top-5,000 downloads in mid-March and about 16% now. Both lost downloads outright while the top 5,000's total grew about a quarter, and the fall holds when we follow a fixed set of models, use a fixed download cut-off instead of the top 5,000, or leave converted copies out.

LLMs and embedding models went from about 45% to about 58%. Converted copies of LLMs and two embedding and reranking repos account for almost all of that gain. Most of the shift happened between mid-April and early June. Speech was about 8% in mid-March and is about 8% now, after a dip to about 6% in May.

Share of top-5,000 downloads by model typeWeekly average from March. Days with counting faults are left out, and so is one April-to-May burst on a single LLM.

LLMsEmbeddingsText classificationVision

Show the numbers: Share of top-5,000 downloads by model type

Share of top-5,000 downloads by model type
Week of	LLMs	Embeddings	Text classification	Vision	Speech and audio	Image and video generation	Time series	Translation and summarisation	Other tagged	Untagged
Mar 9	22.5%	22.2%	13.7%	11.8%	8.8%	1.5%	3.4%	0.9%	1.6%	13.6%
Mar 16	23.0%	22.4%	13.6%	11.9%	8.0%	1.5%	3.4%	0.9%	1.9%	13.5%
Mar 23	23.4%	22.4%	13.7%	12.1%	7.8%	1.4%	3.3%	0.9%	1.7%	13.3%
Mar 30	23.7%	22.8%	13.8%	12.2%	7.9%	1.4%	3.0%	0.9%	1.5%	12.7%
Apr 6	24.1%	23.2%	13.5%	12.4%	7.7%	1.4%	2.9%	0.9%	1.5%	12.4%
Apr 13	25.0%	23.8%	12.9%	12.2%	7.3%	1.4%	2.8%	0.9%	1.5%	12.2%
Apr 20	25.7%	25.1%	12.2%	11.6%	6.9%	1.4%	2.7%	0.9%	1.5%	12.0%
Apr 27	25.8%	26.4%	11.7%	11.1%	6.5%	1.4%	2.8%	1.0%	1.6%	11.8%
May 4	26.0%	27.6%	11.6%	10.2%	6.5%	1.3%	2.8%	1.0%	1.5%	11.5%
May 11	26.6%	28.5%	11.5%	9.3%	6.4%	1.3%	2.7%	1.0%	1.5%	11.3%
May 18	27.3%	28.7%	11.6%	8.4%	6.3%	1.3%	2.7%	0.9%	1.5%	11.3%
May 25	28.1%	28.9%	11.4%	7.7%	6.2%	1.3%	2.6%	0.9%	1.6%	11.3%
Jun 1	28.7%	29.3%	11.1%	7.3%	6.4%	1.3%	2.4%	0.9%	1.7%	11.0%
Jun 8	29.1%	29.5%	10.9%	7.1%	6.4%	1.3%	2.4%	0.9%	1.7%	10.8%
Jun 15	28.9%	29.2%	10.3%	7.1%	8.9%	1.2%	2.3%	1.0%	1.6%	9.4%
Jun 22	28.8%	29.0%	10.3%	7.3%	9.3%	1.2%	2.3%	1.2%	1.6%	9.2%
Jun 29	28.8%	28.9%	10.4%	7.5%	9.2%	1.2%	2.3%	1.3%	1.4%	9.0%
Jul 6	28.8%	29.0%	10.5%	7.5%	8.8%	1.1%	2.3%	1.5%	1.4%	9.2%
Jul 13	28.9%	29.4%	10.7%	7.2%	7.5%	1.2%	2.3%	1.6%	1.3%	9.9%
Jul 20	29.3%	29.2%	11.0%	6.9%	7.2%	1.2%	2.3%	1.6%	1.3%	10.0%
Jul 27	29.4%	29.1%	11.3%	6.5%	7.2%	1.2%	2.6%	1.5%	1.3%	10.0%
Aug 3	29.7%	29.0%	11.6%	6.3%	7.4%	1.1%	2.8%	1.4%	1.2%	9.4%
Aug 10	30.0%	28.7%	11.6%	6.3%	7.4%	1.2%	2.8%	1.4%	1.3%	9.5%
Aug 17	29.4%	28.4%	11.1%	6.2%	7.6%	1.5%	2.8%	1.4%	1.3%	10.2%
Aug 24	29.1%	28.1%	10.4%	6.2%	8.3%	1.6%	2.5%	1.4%	1.4%	10.9%
Aug 31	29.2%	28.0%	9.7%	6.3%	8.7%	1.8%	2.4%	1.4%	1.5%	11.0%
Sep 7	29.3%	28.0%	9.1%	6.3%	8.9%	1.9%	2.5%	1.5%	1.4%	11.1%
Sep 14	29.4%	28.1%	9.0%	6.5%	8.6%	2.0%	2.5%	1.5%	1.3%	11.1%
Sep 21	29.5%	28.1%	9.0%	6.7%	8.6%	2.1%	2.5%	1.4%	1.2%	10.8%
Sep 28	29.8%	28.2%	8.7%	7.0%	8.3%	2.2%	2.4%	1.4%	1.1%	10.8%

Share of downloads by model type: 12 to 13 February, mid-March and the week to 1 October
Model type	12 to 13 Feb	Mid-March	Week to 1 Oct
LLMs	20.6%	22.7%	29.8%
Embeddings and rerankers	21.9%	22.7%	28.1%
Text classification	14.1%	13.5%	8.8%
Speech and audio	7.3%	8.1%	8.4%
Vision	14.8%	11.9%	6.9%
Image and video generation	1.7%	1.5%	2.2%
Time series	2.7%	3.3%	2.4%
Translation and summarisation	1.2%	0.9%	1.4%
Other tagged	2.5%	1.9%	1.2%
Untagged	13.3%	13.5%	10.8%

Model age by type

Now, about 79% of embedding downloads, 76% of text classification and 73% of vision go to repos created in 2023 or earlier, many of them older than March 2022, the earliest creation date Hugging Face records. For LLMs it is 7%. About a third to a half of LLM downloads (47% counting converted copies, 31% without them) and about a third of image and video downloads go to repos created since 12 February, against under 4% for embedding, text classification and vision models.

A new repo can hold an old model: about 60% of the downloads of LLM repos created this year go to converted copies of existing models.

Share of each type's downloads going to repos created in 2023 or earlierWeek to 1 October. Includes repos with Hugging Face's earliest recorded date, which are older than March 2022.

Translation and summarisation98.6%

Embeddings and rerankers79.2%

Text classification76.4%

Vision73.1%

Speech and audio58.8%

Image and video generation15.1%

LLMs7.0%

Time series0.0%

Show the numbers: Share of each type's downloads going to repos created in 2023 or earlier

Share of each type's downloads going to repos created in 2023 or earlier
Model type	2023 or earlier	Created since 12 Feb
Translation and summarisation	98.6%	1.2%
Embeddings and rerankers	79.2%	0.4%
Text classification	76.4%	1.2%
Vision	73.1%	3.7%
Speech and audio	58.8%	10.0%
Image and video generation	15.1%	34.2%
LLMs	7.0%	46.8%
Time series	0.0%	3.2%

Models that move together

Pattern measured, cause inferred

Several of the largest moves happen on the same days across unrelated models, which points to automated or bulk downloading. Hugging Face does not publish who downloads. From about 20 June, t5-small, DETR, SAM, fastText's language identifier and ModernBERT-large all started climbing, and Meta's Prompt-Guard-86M, already rising since May, climbed again; four of the five were back near their earlier levels by early August. About 60 older per-language wav2vec2 speech models have risen and fallen together several times since February.

Models that move in lockstep account for about 36% of speech downloads and 15% of text classification downloads now.

Where fine-tuning shows

Repos created since 2025

Among repos created since 2025, where tags naming the model a repo was built from are common enough to compare, about a third of image and video downloads went to fine-tunes or adapters published by other accounts, against about 4% for LLMs, whose downloaded derivatives are mostly compressed copies. Older model types cannot be compared this way: most of their repos predate those tags. All of these counts cover public repos only; fine-tunes that companies build on their own data stay private and do not appear here.

Turnover at the top

About half the top-100 slots, 53 models, mostly pre-2025 embedding, BERT-class, speech and small Qwen and Llama models, never changed hands in 231 days, and they take about three-quarters of top-100 downloads. The other half turned over: 42 of February's top 100 were out of it by October. Of about 2,900 models created in 2026 that reached the top 5,000, 46 reached the top 100, and two-thirds of those were versions of five LLM launch families: Qwen 3.5, 3.6 and 3.8, Gemma 4 and DeepSeek V4. Embedding, text classification, time series and translation models now get 94 to 98% of their downloads from repos that were already in the top 5,000 in mid-March, against about 57% for LLMs.

About 30% of downloads

LLMs

See the live LLM trends

LLM downloads by model owner

Share of LLM downloads, weekly

Qwen-based models take about 45 to 48% of LLM downloads and have since March. Since February, Google's share more than doubled on Gemma 4 (4% to 10%, after a peak of 14% in July), and Meta's fell from about 9% to 3.5%, with its volume down about 30% since spring. Mistral is under 1%. Credited to the lab whose weights they hold, Chinese labs' models are about three-fifths of LLM downloads, and that share has barely moved since February (about 59% to 61%).

Share of open LLM downloads by base-model ownerWeekly average across the top 5,000 models. Days with counting faults are left out.

QwenGoogleMetaMistral

Show the numbers: Share of open LLM downloads by base-model owner

Share of open LLM downloads by base-model owner
Week of	Qwen	Google	Meta	Mistral	Everyone else
Feb 9	43.4%	4.5%	9.3%	2.0%	40.8%
Feb 16	-	-	-	-	-
Feb 23	-	-	-	-	-
Mar 2	-	-	-	-	-
Mar 9	47.7%	3.5%	8.1%	1.4%	39.2%
Mar 16	47.9%	3.5%	8.2%	1.4%	39.0%
Mar 23	47.8%	3.4%	8.4%	1.4%	39.0%
Mar 30	47.8%	3.5%	8.4%	1.3%	38.9%
Apr 6	47.4%	4.9%	8.1%	1.3%	38.4%
Apr 13	46.5%	6.7%	7.7%	1.2%	37.9%
Apr 20	45.6%	8.4%	7.3%	1.2%	37.5%
Apr 27	44.8%	9.8%	7.0%	1.1%	37.3%
May 4	45.3%	10.5%	7.0%	1.1%	36.1%
May 11	46.1%	10.9%	6.7%	1.1%	35.2%
May 18	45.8%	11.0%	6.5%	1.1%	35.6%
May 25	45.8%	11.3%	6.0%	1.0%	35.9%
Jun 1	45.9%	11.4%	5.7%	0.9%	36.1%
Jun 8	45.8%	11.7%	5.6%	0.8%	36.1%
Jun 15	45.4%	13.0%	5.3%	0.7%	35.5%
Jun 22	45.5%	13.6%	5.2%	0.7%	35.0%
Jun 29	44.9%	14.2%	5.3%	0.6%	35.0%
Jul 6	44.5%	14.3%	5.1%	0.6%	35.5%
Jul 13	44.0%	14.3%	5.0%	0.6%	36.1%
Jul 20	44.1%	13.7%	4.9%	0.6%	36.8%
Jul 27	44.5%	13.2%	4.7%	0.6%	37.1%
Aug 3	44.7%	12.9%	4.5%	0.6%	37.2%
Aug 10	45.6%	12.3%	4.4%	0.6%	37.1%
Aug 17	46.1%	12.0%	4.1%	0.6%	37.1%
Aug 24	46.8%	11.9%	3.7%	0.6%	36.9%
Aug 31	48.0%	10.9%	3.4%	0.6%	37.1%
Sep 7	47.9%	10.4%	3.3%	0.6%	37.8%
Sep 14	47.2%	10.5%	3.4%	0.6%	38.3%
Sep 21	47.7%	10.3%	3.5%	0.6%	37.9%
Sep 28	47.2%	10.3%	3.6%	0.6%	38.3%

Converted copies of LLMs

LLMs went from about 23% of downloads in mid-March to about 30% by June, and have sat near 29 to 30% since. In spring, converted copies and original repos both grew; since June, copies kept growing while original repos fell back, so by October the gain since March sits almost entirely in converted copies (about 85 to 100%, depending on whether labs' own compressed releases count as copies). GGUF, compressed lower-precision and MLX versions (MLX is Apple's framework for running models on Macs) grew from about 5% to about 12% of all downloads, while LLM repos that are not copies are no higher than in March. Part of the late rise lines up with a platform-wide dip and rebound in mid-August, so some of it may come from how downloads are counted.

LLM downloads: converted copies and everything elseShare of all top-5,000 downloads, monthly average, with one April-to-May burst left out.

Other LLM reposConverted copies

Show the numbers: LLM downloads: converted copies and everything else

LLM downloads: converted copies and everything else
Month	Other LLM repos	Converted copies
Mar 2026	18.0%	5.2%
Apr 2026	19.1%	5.7%
May 2026	20.0%	6.9%
Jun 2026	20.9%	7.9%
Jul 2026	20.1%	8.9%
Aug 2026	19.3%	10.2%
Sep 2026	17.2%	12.3%

LLM licences

Permissively licensed models, mostly Apache-2.0, went from 69% to 78% of LLM downloads since mid-March, and the shift holds on a fixed set of models. Qwen and now Gemma 4 ship under Apache-2.0, while Llama-licensed models stalled. Across model types, time series (about 97%) and embedding models (about 95%) are the most permissive, and image and video models the least: under half carry a permissive licence and about half carry custom ones.

LLM sizes

Four of 2026's most-downloaded new open LLM lines are 26 to 31B parameters: Qwen 3.6 and 3.8 27B, and Gemma 4 26B and 31B. Those four model lines took the 10 to 35B size class from about a fifth to about a third of LLM downloads. Without them, that size class is flat.

Copies, fine-tunes and big-lab releases

Among popular derivatives of open LLMs, 70 to 90% are compressed copies of a lab's model, which appear within about a week of a release; fine-tuned versions follow about two weeks later. Unsloth, which makes an open-source fine-tuning library, published the first popular derivative for 18 of 34 major 2026 releases. Official releases from big labs were about twice as likely as community copies to still be at or above their launch-month downloads three months later, in a sample of 18 to 32 releases, depending on how launch dates are counted.

About 28% of downloads

Embedding and reranker models

See the live embedding and reranker trends

Retrieval models at the top

all-MiniLM-L6-v2 from sentence-transformers was the most downloaded model of any type on all 231 days, with about 243M downloads in the 30 days to 1 October, and five of the ten most-downloaded models are retrieval models. Embedding and reranker models take about 28% of downloads from about 8% of the top-5,000 slots, and three repos hold half of that. Their counts include automated pulls by tools and pipelines.

Rerankers

Cause not shown

Rerankers went from about 1.6% to 5.1% of all downloads since mid-March, or from about 7% to 18% of the type. cross-encoder/ms-marco-MiniLM-L6-v2 is about 60% of that. It and two models from BAAI, the Beijing Academy of Artificial Intelligence, bge-small-en-v1.5 and bge-reranker-v2-m3, climbed together from mid-April to late June and are about where they were then, and their day-to-day changes track each other more closely than any other pair among models averaging over 3M downloads a month. That pattern fits one large consumer, such as a framework's default models; the data cannot show who it is. Without its two biggest gainers, ms-marco-MiniLM-L6-v2 and bge-small-en-v1.5, the type's gain since mid-March is about 2 points, against 5 with them.

Rerankers’ share of all model downloadsWeekly mean of daily shares, March to 1 October. Test fixtures and faulty or stale days excluded; first and last weeks are partial.

Rerankers

Show the numbers: Rerankers’ share of all model downloads

Rerankers’ share of all model downloads
Week of	Rerankers
Mar 9	1.5%
Mar 16	1.6%
Mar 23	1.7%
Mar 30	1.8%
Apr 6	2.0%
Apr 13	2.2%
Apr 20	2.4%
Apr 27	2.6%
May 4	3.0%
May 11	3.3%
May 18	3.7%
May 25	4.1%
Jun 1	4.4%
Jun 8	4.6%
Jun 15	4.8%
Jun 22	4.8%
Jun 29	4.9%
Jul 6	5.0%
Jul 13	5.1%
Jul 20	5.2%
Jul 27	5.2%
Aug 3	5.1%
Aug 10	5.1%
Aug 17	5.0%
Aug 24	4.9%
Aug 31	4.9%
Sep 7	4.9%
Sep 14	5.0%
Sep 21	5.0%
Sep 28	5.1%

New embedding models

Repos created in 2025 or 2026 hold about 7% of embedding and reranker downloads; the largest, Qwen3-Embedding-0.6B, about 1%. The shift that did happen is between established models: BAAI went from about 10% to 21% of the type and sentence-transformers from about 62% to 46%, with copies left out, on models uploaded in 2024 or earlier.

About 8% of downloads

Speech and audio models

See the live speech and audio trends

Converted copies of speech models

Counting converted copies, speech and audio held about 8% of downloads; without them it fell from about 7.6% to 6.7% since mid-March. Converted copies, such as faster-whisper, WhisperKit, the audio.cpp and transcribe.cpp GGUF builds, and MLX versions, went from about 6% to 20% of the type's downloads, and five uploaders hold about 87% of them. OpenAI's own Whisper repos held steady. Converted copies of NVIDIA's speech models more than doubled their share, and Whisper still has about four times their downloads.

Speech downloads: copies and other reposShare of all model downloads, monthly mean. March starts on 13 March; test fixtures and faulty or stale days excluded.

Other speech reposConverted copies

Show the numbers: Speech downloads: copies and other repos

Speech downloads: copies and other repos
Month	Other speech repos	Converted copies
Mar 2026	7.5%	0.5%
Apr 2026	6.6%	0.7%
May 2026	5.3%	0.8%
Jun 2026	7.1%	0.8%
Jul 2026	7.0%	0.9%
Aug 2026	6.4%	1.3%
Sep 2026	7.0%	1.6%

The June speech jump

Cause not shown

Speech recognition's jump in June came from 60 to 90 older per-language wav2vec2 models rising together. It began in the same days as a Hugging Face counting fault and recurred in late August and September without one. Without these pulses, speech's share is flat at about 6 to 7%.

pyannote

All repos from pyannote, an open-source speaker diarization toolkit, fell from about 2.2% to 1.0% of downloads since mid-March in a steady slide, their downloads down about 40% while the market grew about a quarter. Its newer community-1 diarization model rose from about 1.4M to 5.5M downloads a month, a fraction of the older pipeline's losses.

About 7% of downloads

Vision models

See the live vision trends

Where vision's share went

Vision models fell from about 12% to 7% of downloads since mid-March. Image classification accounts for about 85% of the lost downloads, and vision still fell on a fixed set of models. Classifiers whose names say they detect explicit content, age, gender or ethnicity fell from about 47% of image classification downloads to about 10%; Falconsai's NSFW detector alone went from about 42M downloads a month in mid-March to 2.4M now, mostly in a drain from 19 April. The twelve biggest losers lost about 83M downloads a month, while the other vision repos together grew about 11% as the market grew about a quarter.

How vision’s task mix changedShare of vision downloads. Mid-March versus the week to 1 October; converted copies, test fixtures and junk repos excluded. This compares composition, not total volume.

Show the numbers: How vision’s task mix changed

How vision’s task mix changed
Group	Mid-March	Week to 1 Oct
Image classification	46.7%	31.2%
Zero-shot image classification	25.7%	36.8%
Object detection	3.8%	3.9%
Image features	6.4%	10.2%
Other tasks	17.5%	18.0%

Zero-shot models and Google

Zero-shot image classification, the CLIP and SigLIP family, is now the largest vision task at about 37% of the type, because image classification shrank. Google's share of vision downloads roughly doubled since mid-March, from about 6% to between 12% and 16% depending on the week, on SigLIP 2 and ViT, helped by vision shrinking.

New vision models

The newcomers that count are Google's SigLIP 2, at about 12M downloads a month against about 4M in mid-March (about 8M without a mid-September spike on one repo), and Meta's DINOv3 and SAM 3, at about 2M to 4M each; the older DINOv2 grew more than DINOv3. The detectors RF-DETR and D-FINE barely reach the top 5,000.

About 9% of downloads

Text classification models

As on our text classification page, this group includes base encoders such as BERT (about half its downloads), entity taggers and classifiers.

See the live text classification trends

BERT's share

Text classification fell from about 13.5% to 8.8% of downloads since mid-March. The type's own downloads are about where they were in February while the market grew, and bert-base-uncased and roberta-large account for about half of the lost share. The original BERT checkpoint's share of downloads is at its lowest in our data, while models built on the BERT design, mostly embedding models, grew to about a quarter of all downloads.

What contributed to text classification’s lost shareDrop in share of all model downloads from mid-March to the week to 1 October, in percentage points. Other repos is the net of gains and losses; test fixtures excluded.

Show the numbers: What contributed to text classification’s lost share

What contributed to text classification’s lost share
Group	Share lost
BERT base uncased	1.70 pp
RoBERTa large	0.77 pp
Other repos (net)	2.24 pp

OpenMed's medical models

OpenMed, which publishes medical-text entity models on Hugging Face, has roughly 350 models in the top 5,000: about half of all text classification models there and about 15% of the type's downloads. About 70% of them sit within 5% of 95,000 downloads a month, and their daily counts move together more closely than other classifiers'; the data cannot show why.

Newer encoders and guardrail classifiers

Base encoders created since 2024 are about 5% of base-encoder downloads, flat since spring, and ModernBERT-base swings between about 1M and 10M downloads a month. Guardrail classifiers grew from about 0.7M to 5.3M downloads a month; Meta's Prompt-Guard-86M is about three-quarters of that and rose in two steps, in May and late June, the same weeks as several unrelated older models.

About 2% of downloads by task tag, about 5% with ComfyUI repackages

Image and video generation models

See the live image and video trends

ComfyUI repackages

Repos tagged for image or video generation are about 2.2% of downloads. Single-file weights repackaged for ComfyUI, an open-source node-based app for running image and video models, carry no task tag; counting them, the group is about 5.3%, and the repackages take about half of its downloads. For several recent models, including Krea-2, Z-Image, Wan, MiniMax-H3 and Qwen-Image, the repackage was downloaded several times more than the lab's own repo. Hugging Face counts the two kinds of repo differently, so this shows direction rather than a ratio of use.

Image and video beyond the task tagsShare of all model downloads; mid-March versus the week to 1 October. Tagged models exclude 3D. Untagged uses the narrow ComfyUI-style repackage definition. Different file-counting rules mean this is not a comparison of usage.

Show the numbers: Image and video beyond the task tags

Image and video beyond the task tags
Group	Mid-March	Week to 1 Oct
Tagged image / video	1.5%	2.2%
Untagged repackages	1.8%	3.0%

Turnover and licences

MiniMax-H3, uploaded on 28 July, and its repackages and fine-tunes are about a quarter of the group's downloads, while Wan models fell from about 43% to 11%. Under half of image and video downloads carry a permissive licence, against about four-fifths for LLMs; about half carry custom licences such as OpenRAIL or Hugging Face's catch-all "other", among them FLUX.1-dev, MiniMax-H3 and LTX.

About 4% of downloads together

Time series and translation models

See the live time series trends See the live translation trends

Chronos and forecasting

Amazon's Chronos models, uploaded under the amazon and autogluon accounts, are about 88% of forecasting downloads, and Chronos-2 alone about half. Forecasting models were about 2.4% of downloads in the week to 1 October, down from about 3.3% in March and roughly flat since May, so their downloads did not grow while the market did. Google's TimesFM appears to jump on 9 September, which looks like a change in counting.

Chronos’ share of forecasting downloadsShare of forecasting downloads, mid-March versus the week to 1 October. Uses page rules excluding tokenizers and test fixtures; Amazon and Autogluon releases are grouped together.

Show the numbers: Chronos’ share of forecasting downloads

Chronos’ share of forecasting downloads
Group	Mid-March	Week to 1 Oct
Chronos-2	36.9%	54.6%
Other Chronos	49.8%	33.7%
Other forecasting	13.3%	11.7%

Translation and summarisation

Models tagged for translation or summarisation went from about 0.9% to 1.4% of downloads, and one repo explains the rise: google-t5/t5-small, a general text-to-text model that Hugging Face tags as translation, went from about 1.8M downloads a month in mid-March to 24M now, almost all from a climb that began around 20 June, the same week as the group of unrelated models above. Without it the share fell from about 0.8% to 0.55%. Tagged summarisation halved, and no translation model uploaded since 2025 has become large; Google's TranslateGemma is tagged as an LLM.

Translation and summarisation without T5-smallShare of all model downloads, mid-March versus the week to 1 October. Both comparisons use page rules with converted copies excluded; T5-small is tagged as translation.

Show the numbers: Translation and summarisation without T5-small

Translation and summarisation without T5-small
Group	Mid-March	Week to 1 Oct
Including T5-small	0.9%	1.4%
Without T5-small	0.8%	0.5%

How to read Hugging Face download numbers

Each of these would have produced a confident, wrong conclusion. Anyone using Hugging Face download numbers runs into the same ones.

How to read these numbers. Hugging Face's documentation says every request for a model's config file counts as a download, including requests that only check whether the file has changed, and that each file counts separately in GGUF repos (GGUF is a file format for running models locally) and in single-file image repos. Nothing is deduplicated and automated traffic is counted. These are Hugging Face numbers only: models that also ship through other channels, such as their own packages, are undercounted here. We compare shares, which hold up better than totals, and name the window each time: "mid-March" is the week of 15 to 21 March, and "now" is the week to 1 October.

The June counting faultTotal 30-day downloads across the top 5,000 models, billions

Total 30-day downloadsCounting fault

Show the numbers: The June counting fault

The June counting fault
Day	Total 30-day downloads
Mar 13	2.17B
Mar 14	2.17B
Mar 15	2.18B
Mar 16	2.17B
Mar 17	2.17B
Mar 18	2.18B
Mar 19	2.19B
Mar 20	2.21B
Mar 21	2.19B
Mar 22	2.19B
Mar 23	2.18B
Mar 24	2.17B
Mar 25	2.18B
Mar 26	2.18B
Mar 27	2.18B
Mar 28	2.18B
Mar 29	2.19B
Mar 30	2.18B
Mar 31	2.18B
Apr 1	2.19B
Apr 2	2.13B
Apr 3	2.13B
Apr 4	2.14B
Apr 5	2.13B
Apr 6	2.12B
Apr 7	2.12B
Apr 8	2.14B
Apr 9	2.14B
Apr 10	2.14B
Apr 11	2.14B
Apr 12	2.16B
Apr 13	2.15B
Apr 14	2.16B
Apr 15	2.17B
Apr 16	2.18B
Apr 17	2.20B
Apr 18	2.22B
Apr 19	2.24B
Apr 20	2.26B
Apr 21	2.28B
Apr 22	2.31B
Apr 23	2.34B
Apr 24	2.36B
Apr 25	2.38B
Apr 26	2.40B
Apr 27	2.42B
Apr 28	2.44B
Apr 29	2.46B
Apr 30	2.47B
May 1	2.58B
May 2	2.60B
May 3	2.61B
May 4	2.61B
May 5	2.62B
May 6	2.64B
May 7	2.66B
May 8	2.68B
May 9	2.69B
May 10	2.70B
May 11	2.69B
May 12	2.69B
May 13	2.70B
May 14	2.73B
May 15	2.74B
May 16	2.74B
May 17	2.73B
May 18	2.72B
May 19	2.71B
May 20	2.72B
May 21	2.72B
May 22	2.72B
May 23	2.71B
May 24	2.71B
May 25	2.69B
May 26	2.69B
May 27	2.69B
May 28	2.69B
May 29	2.68B
May 30	2.67B
May 31	2.68B
Jun 1	2.68B
Jun 2	2.67B
Jun 3	2.69B
Jun 4	2.69B
Jun 5	2.69B
Jun 6	2.68B
Jun 7	2.68B
Jun 8	2.67B
Jun 9	2.67B
Jun 10	2.46B
Jun 11	2.44B
Jun 12	2.43B
Jun 13	2.41B
Jun 14	1.80B
Jun 15	1.81B
Jun 16	1.84B
Jun 17	1.88B
Jun 18	2.04B
Jun 19	2.36B
Jun 20	2.69B
Jun 21	2.69B
Jun 22	-
Jun 23	2.74B
Jun 24	2.75B
Jun 25	2.77B
Jun 26	2.73B
Jun 27	2.74B
Jun 28	2.74B
Jun 29	2.73B
Jun 30	2.72B
Jul 1	2.71B
Jul 2	2.73B
Jul 3	2.74B
Jul 4	2.73B
Jul 5	2.73B
Jul 6	2.73B
Jul 7	2.77B
Jul 8	2.79B
Jul 9	2.85B
Jul 10	2.90B
Jul 11	2.89B
Jul 12	2.85B
Jul 13	2.83B
Jul 14	2.83B
Jul 15	2.88B
Jul 16	2.86B
Jul 17	2.81B
Jul 18	2.86B
Jul 19	2.86B
Jul 20	2.80B
Jul 21	2.72B
Jul 22	2.74B
Jul 23	2.85B
Jul 24	2.89B
Jul 25	2.90B
Jul 26	2.90B
Jul 27	2.89B
Jul 28	2.89B
Jul 29	2.91B
Jul 30	2.91B
Jul 31	2.93B
Aug 1	2.94B
Aug 2	2.90B
Aug 3	2.97B
Aug 4	2.98B
Aug 5	3.01B
Aug 6	2.99B
Aug 7	2.91B
Aug 8	2.87B
Aug 9	2.83B
Aug 10	2.80B
Aug 11	2.77B
Aug 12	2.75B
Aug 13	2.75B
Aug 14	2.88B
Aug 15	2.94B
Aug 16	3.00B
Aug 17	2.98B
Aug 18	2.98B
Aug 19	2.99B
Aug 20	2.98B
Aug 21	2.97B
Aug 22	2.97B
Aug 23	2.96B
Aug 24	2.94B
Aug 25	2.94B
Aug 26	2.96B
Aug 27	2.96B
Aug 28	2.90B
Aug 29	2.78B
Aug 30	2.85B
Aug 31	2.90B
Sep 1	2.88B
Sep 2	2.90B
Sep 3	2.87B
Sep 4	2.81B
Sep 5	2.90B
Sep 6	2.91B
Sep 7	2.88B
Sep 8	2.86B
Sep 9	2.88B
Sep 10	2.88B
Sep 11	2.88B
Sep 12	2.88B
Sep 13	2.88B
Sep 14	2.86B
Sep 15	2.85B
Sep 16	2.86B
Sep 17	2.88B
Sep 18	2.87B
Sep 19	2.87B
Sep 20	2.86B
Sep 21	2.84B
Sep 22	2.82B
Sep 23	2.83B
Sep 24	2.81B
Sep 25	2.81B
Sep 26	2.76B
Sep 27	2.76B
Sep 28	2.74B
Sep 29	2.73B
Sep 30	2.73B
Oct 1	2.73B

Between 10 and 19 June, almost every model's reported 30-day downloads fell by about a third within a few days, from 2.67B to 1.80B in total, then recovered by 20 June while likes stayed put. A smaller platform-wide dip and rebound followed in August: the total slid from 3.01B on 5 August to 2.75B on 12 August and was back to 3.00B by 16 August, with converted LLM copies, OpenMed and speech conversions rebounding the most. Comparisons that span those dates need care.

The August GGUF stepGGUF share of open LLM downloads, daily

GGUF share of LLM downloadsSynchronised step

Show the numbers: The August GGUF step

The August GGUF step
Day	GGUF share of LLM downloads
Mar 13	8.1%
Mar 14	8.1%
Mar 15	8.2%
Mar 16	8.4%
Mar 17	8.5%
Mar 18	8.4%
Mar 19	8.4%
Mar 20	8.3%
Mar 21	8.5%
Mar 22	8.5%
Mar 23	8.7%
Mar 24	8.8%
Mar 25	8.8%
Mar 26	8.9%
Mar 27	8.9%
Mar 28	9.0%
Mar 29	9.1%
Mar 30	9.3%
Mar 31	9.4%
Apr 1	9.3%
Apr 2	9.3%
Apr 3	9.2%
Apr 4	9.1%
Apr 5	9.2%
Apr 6	9.2%
Apr 7	9.2%
Apr 8	9.3%
Apr 9	9.4%
Apr 10	9.6%
Apr 11	9.5%
Apr 12	9.7%
Apr 13	9.9%
Apr 14	9.8%
Apr 15	9.8%
Apr 16	9.9%
Apr 17	9.8%
Apr 18	9.6%
Apr 19	9.5%
Apr 20	9.5%
Apr 21	9.3%
Apr 22	9.3%
Apr 23	9.1%
Apr 24	9.0%
Apr 25	9.0%
Apr 26	8.9%
Apr 27	8.8%
Apr 28	8.6%
Apr 29	8.5%
Apr 30	8.4%
May 1	8.3%
May 2	8.3%
May 3	8.3%
May 4	8.2%
May 5	8.2%
May 6	8.2%
May 7	8.1%
May 8	8.0%
May 9	7.9%
May 10	7.8%
May 11	7.8%
May 12	7.7%
May 13	7.7%
May 14	7.6%
May 15	7.6%
May 16	7.7%
May 17	7.9%
May 18	8.0%
May 19	8.1%
May 20	8.1%
May 21	8.1%
May 22	8.2%
May 23	8.2%
May 24	8.3%
May 25	8.5%
May 26	8.6%
May 27	8.7%
May 28	8.8%
May 29	8.9%
May 30	9.0%
May 31	9.0%
Jun 1	9.1%
Jun 2	9.1%
Jun 3	9.0%
Jun 4	8.9%
Jun 5	8.9%
Jun 6	8.9%
Jun 7	9.0%
Jun 8	9.2%
Jun 9	9.2%
Jun 10	9.3%
Jun 11	9.4%
Jun 12	9.4%
Jun 13	9.1%
Jun 14	9.9%
Jun 15	10.0%
Jun 16	10.0%
Jun 17	9.9%
Jun 18	9.5%
Jun 19	9.6%
Jun 20	9.5%
Jun 21	9.6%
Jun 22	-
Jun 23	9.7%
Jun 24	9.7%
Jun 25	9.7%
Jun 26	9.7%
Jun 27	9.8%
Jun 28	9.9%
Jun 29	9.9%
Jun 30	9.9%
Jul 1	9.9%
Jul 2	10.0%
Jul 3	10.0%
Jul 4	10.1%
Jul 5	10.2%
Jul 6	10.0%
Jul 7	9.9%
Jul 8	9.9%
Jul 9	9.9%
Jul 10	9.8%
Jul 11	9.8%
Jul 12	9.8%
Jul 13	9.8%
Jul 14	9.8%
Jul 15	9.6%
Jul 16	9.6%
Jul 17	9.6%
Jul 18	9.5%
Jul 19	9.5%
Jul 20	9.4%
Jul 21	9.4%
Jul 22	9.3%
Jul 23	9.1%
Jul 24	9.2%
Jul 25	9.2%
Jul 26	9.1%
Jul 27	9.2%
Jul 28	9.2%
Jul 29	9.2%
Jul 30	9.2%
Jul 31	9.4%
Aug 1	9.3%
Aug 2	9.3%
Aug 3	9.4%
Aug 4	9.5%
Aug 5	9.5%
Aug 6	9.5%
Aug 7	9.6%
Aug 8	9.8%
Aug 9	10.0%
Aug 10	10.2%
Aug 11	10.3%
Aug 12	10.6%
Aug 13	10.6%
Aug 14	11.5%
Aug 15	12.0%
Aug 16	12.6%
Aug 17	13.1%
Aug 18	13.6%
Aug 19	13.9%
Aug 20	14.4%
Aug 21	14.9%
Aug 22	15.3%
Aug 23	15.9%
Aug 24	16.4%
Aug 25	16.9%
Aug 26	17.2%
Aug 27	17.6%
Aug 28	17.9%
Aug 29	18.2%
Aug 30	18.6%
Aug 31	19.2%
Sep 1	19.6%
Sep 2	19.8%
Sep 3	19.9%
Sep 4	20.5%
Sep 5	20.7%
Sep 6	20.8%
Sep 7	21.1%
Sep 8	21.3%
Sep 9	21.4%
Sep 10	21.4%
Sep 11	21.6%
Sep 12	21.8%
Sep 13	21.9%
Sep 14	22.1%
Sep 15	22.1%
Sep 16	22.0%
Sep 17	21.9%
Sep 18	22.0%
Sep 19	22.1%
Sep 20	22.2%
Sep 21	22.4%
Sep 22	22.4%
Sep 23	22.3%
Sep 24	22.4%
Sep 25	22.4%
Sep 26	22.5%
Sep 27	22.6%
Sep 28	22.7%
Sep 29	22.8%
Sep 30	22.9%
Oct 1	23.0%

GGUF, the file format used to run models locally, held at about 8 to 10% of LLM downloads until mid-August, then rose from about 10.6% on 12 August to about 23% on 1 October. Existing GGUF repos from most large publishers rose together, about 1.56 times between 10 August and 12 September while the top 5,000's total barely moved, starting with the mid-August rebound. Hugging Face also counts each GGUF file separately. We read the jump as unexplained rather than as proof of a shift to running models locally.

Models moving in step. Unrelated models in different types start and stop climbing on the same days, in late April, late June and mid-August. Before calling a model the fastest-growing, check what else moved that week.
Downloads by file. Most models count one config-file request per download; GGUF and single-file image weights count each file. Levels are hard to compare across model types.
Repos with no task tag. About 11% of downloads go to repos with no task tag: ComfyUI repackages, Google's ELECTRA encoder (about 46M downloads a month, more than bert-base-uncased), YOLO detectors and speaker models. For image and video generation, the untagged part is bigger than the tagged part.
New repos, old models. About 60% of the downloads of LLM repos created this year go to converted copies, and a re-upload gets a new creation date.
Placeholder creation dates. About 540 older repos carry Hugging Face's 2022-03-02 placeholder date as their creation date: about 32% of all downloads, 59% of embedding downloads and 69% of text classification. Their real age is older than that date.
Changing task tags. GLM-OCR moved from vision to LLM and back; ComfyUI's Wan 2.2 repackage gained a video tag on 17 August and would have topped the image and video ranking overnight.
Slot counts. OpenMed holds about 350 of the 5,000 slots and test-fixture repos about 220, so shares of downloads say more than model counts.
Frozen days. On 14 February and 14 March our collector stored a copy of the previous day's counts, and about 12 more days were partly frozen on Hugging Face's side, 10 of them Thursdays. They affect LLMs more than other types.
Renames. 95 repos changed names, 36 of them between organisations, and took their download counts with them.
Single-source bursts. For about two weeks from mid-April, a burst consistent with a single source added roughly 13 million downloads a day to Qwen3-VL-2B-Instruct, briefly making it 7% of all top-5,000 downloads and shrinking every other type's share that month.
Single-repo collapses. The Falconsai NSFW detector lost about 1M downloads a day from 19 April, a drain consistent with one large consumer switching off, and on its own accounts for about two-fifths of vision's fall.
A missing day. 22 June has no data.

About this data

We record the top 5,000 models on Hugging Face by 30-day downloads every day, starting 12 February 2026; from 14 February to 12 March we kept the top 1,000 only, and 22 June is missing. That is 1,053,899 rows covering 13,182 models up to 1 October. Model types follow our trends pages: each repo is placed by its Hugging Face task tag, and format conversions and re-uploads are identified by format words, converted-weights libraries and base-model tags. Every finding here was recomputed with independent code and then challenged by reviewers whose job was to disprove it.

Azumo collected and stored daily snapshots of download counts and model metadata from the Hugging Face API. This report analyzes that historical collection for 12 February to 1 October 2026. The models are published by their respective developers and hosted on Hugging Face; the collection and analysis are Azumo’s original research.

Source documentation: how Hugging Face counts downloads →