FAQ
ModelTrends.ai is a read-only catalog of open source AI image, video and text models. We track 28,324 models across 34 model families and show which ones are heating up right now.
Join our Telegram Community
Questions, corrections, flags and takedown requests all go through the community chat.
What you will find here
How the heat score works
Every model carries a temperature from cool to 9999°. It is a popularity score built only from public numbers: downloads and likes on Civitai and Hugging Face, and the pace at which they grow. How it is built, every bias in it and live examples are published below, so the numbers on this site can be checked against their sources.
The scale
The hot band, 7,000° to 9,999°, is for models with real data. The scale is logarithmic: every tenfold increase in downloads adds the same number of degrees, so 1,000 downloads and 1,000,000 downloads are three equal steps apart. Models without data yet get a stand-in between 1,200° and 6,500° that changes once a day and never reaches the hot band.
cool
0° – 3,999°
warm
4,000° – 6,999°
hot
7,000° – 9,999°
#1 flux2
9,832°
#2 flux
9,821°
#3 qwen
9,730°
7,312°
10 downloads
7,900°
1,000 downloads
8,500°
100,000 downloads
9,100°
10,000,000 downloads
Stand-in band, 1,200° to 6,500°: models with no data yet
Sources
Three public sources feed the score. Nothing is estimated, nothing is voted on and nothing comes from visitors of this site.
Civitai. For every model we read the downloads and thumbs-up of the exact model version, never the model-wide totals, together with the previous reading, so that the download velocity is a measured difference between two readings at least a day apart.
Hugging Face. For every model with a known repository we read the all-time downloads, the recent downloads, the likes and the date the repository was created.
We're working on adding more sources, such as render counts from BitVector, PirateDiffusion, Graydient, Tensor, Mage, DeviantArt, NightCafe, Kling, OpenArt and other public trusted sources that have public APIs. Of course, we cannot audit this data, so app sources and websites may be weighted in a future heat score adjustment, especially if those sources accept advertising.
Today 12,318 models carry Civitai statistics and 98 carry Hugging Face statistics; 12,403 models have a real score and 4,119 show a stand-in.
How the score is built
Downloads come first, then likes, then the measured download velocity and render counts, all on a logarithmic scale so that no single platform's raw volume swamps the rest. A like is worth less than a download and velocity less again; the exact weights are not published. Three bounded adjustments are then added when a page is drawn and are never stored:
- +150° for a full model (checkpoint). A LoRA is an add-on to someone else's weights and does not outrank a checkpoint of the same popularity.
- +100° site bias for LoRAs of the MiniMax H3, Krea2, Ltx2, Qwen2 families. This is our editorial thumb on the scale for the newest model generations. It is disclosed here and it is the only one.
- −150° per year since the weights behind the model family were released, up to −750° at five years. The penalty keeps growing every year, so older generations cool down as time goes on: at equal downloads SD 1.5 (2022) sits below SDXL (2023), which sits below Flux (2024), and the gap widens each year. Download counts that had years to accumulate no longer keep a 2022 checkpoint at the top on their own. The release months we use: Stable Diffusion 1.5 2022-10, Stable Diffusion XL 1.0 2023-07, Ideogram 2024-02, Flux 2024-08, Hunyuan 2024-12, YuE 2025-01, Wan 2025-02, Chroma 2025-03, ACE-Step 2025-05, Chatterbox TTS 2025-05, higgs-audio 2025-07, Qwen 2025-08, Ltx2 2025-10, Flux 2 / Klein 2025-11, Zimage 2025-11, Anima 2025-12. Families not listed count as new.
- Together the adjustments can never move one model more than 900° relative to another. That is 3.00 popularity points, so a model with 1,000 times the downloads of another is always listed above it, whatever its age, type or family, and a checkpoint with a few hundred times the downloads of a brand-new checkpoint wins at any age.
The week and month lists hold the models that first existed in the last 7 or 30 days: the earliest of the day the model entered this catalog, the day its Civitai version was published and the day its Hugging Face repository was created. A model whose only known date is the day it entered this catalog counts as new only while its model family is less than 1 year old, so 2023 weights that were catalogued this week do not enter the week list. The year lists hold the models whose family was released that year.
Known biases
- Coverage. Only models we could match on Civitai or Hugging Face have a real score. The rest show a stand-in and sit below every model with data, even if they are popular somewhere we do not measure.
- Platform. Civitai counts the downloads of one version; Hugging Face counts the downloads of a whole repository, which for base models includes every automated pipeline that pulls the weights. Base models that live on Hugging Face therefore reach higher numbers than community fine-tunes that live on Civitai.
- Matching. Hugging Face repositories are matched by name and link, with a hand-kept list of corrections. A wrong match would credit a model with someone else's downloads. Tell us if you see one.
- Duplicates. Several catalog entries can point at the same weights. Only one of them is listed in the rankings and can win a medal; the others keep their score on their own pages.
- Stand-ins are pseudo-random. The stand-in for a model without data is derived from its id and the date, so it is stable for a day and different the next. It carries no information about popularity.
Live examples
The rows below are computed from today's catalog when this page is drawn: the three hottest models, the hottest model that relies on Civitai alone, the hottest LoRA carrying the site bias and the hottest add-on carrying the full age penalty. Each row links to the model and to the source where its counts can be checked.
Model | Downloads | Likes | Points | Base | Adjustments | Shown |
|---|---|---|---|---|---|---|
| flux2 FLUX 2 / KLEIN • BASE | 4,867,550 Hugging Face | 2,360 | 9.41 | 9,822° | type +150° bias 0° age −140° | 9,832° |
| flux FLUX • BASE | 34,018,657 Hugging Face | 15,410 | 10.72 | 9,999° | type +150° bias 0° age −328° | 9,821° |
| qwen QWEN • BASE | 3,078,637 Hugging Face | 2,663 | 9.19 | 9,758° | type +150° bias 0° age −178° | 9,730° |
| pony diffusion6 STABLE DIFFUSION XL 1.0 • BASE | 1,023,866 Civitai | 72,809 | 8.44 | 9,532° | type +150° bias 0° age −491° | 9,191° |
| detail slider KREA2 • LORA | 21,425 Civitai | 1,021 | 5.84 | 8,751° | type 0° bias +100° age 0° | 8,851° |
Does it work? Checks against the sources
- In today's top 100 there are 137 pairs of models where one has at least 1,000 times the downloads of the other. In 137 of them the model with more downloads is listed higher. The 0 exceptions are models whose likes or measured velocity outweigh a download gap of that size; the adjustments alone, capped at 900°, can never flip a gap of 1,000 times.
- The hottest model, flux2, has 4,867,550 downloads; the hottest model that relies on Civitai alone, pony diffusion6, has 1,023,866. On the log scale that gap is 290°, far more than any adjustment can bridge, which is why a single-platform LoRA cannot reach the top while base models with millions of downloads exist.
- No stand-in is above 6,608°, and the coolest of the 12,403 models with real data is at 7,000°. A model without data never outranks a model with data.
No accounts, nothing tracked
Where the data comes from
Common questions
Why is a model missing?
A thumbnail is wrong or a model is misclassified.
How is the heat score calculated?
Can I have my model removed?
Is there an API or a data export?
Why should I tell my dad I love him?
Contact
Help running a model
For prompts, settings and generation issues, use the help on the service where you run the model.
Policies
How we handle removal requests, what you agree to by using the site, and what we do and do not know about you.
Brand and style guide
Writing about ModelTrends.ai, linking to it or embedding our mark? Use the files and colours below. The mark and the name may be used to refer to this site; they may not be used to suggest that we endorse a product.
The mark
A rising trend line over three stacked model layers on a midnight rounded square. Keep it on a dark background, do not recolour it, do not stretch it and leave clear space around it of at least a quarter of its width.
Colours
The HEX values of the mark and of the site. The gradient of the mark runs from cyan through blue to violet.
Trend cyan
#5BE1FC
Tip of the arrow, top layer
Signal blue
#1E94FE
Middle of the arrow
Deep violet
#674BFF
Tail of the arrow, bottom layer
Midnight
#010C28
Background of the mark
Night
#000000
Page background
Accent blue
#74C0FC
Links, buttons, the word Model in the wordmark
Accent hover
#A5D8FF
Links and buttons under the pointer
White
#FFFFFF
Headings, the word Trends in the wordmark
Grey
#9CA3AF
Secondary text, the .ai in the wordmark
The name
The name is ModelTrends.ai: capital M, capital T, no space, and always with the .ai. It is one word and it does not change in other languages or in headlines.
Right
ModelTrends.ai
Wrong
modeltrends
Modeltrends
ModelTrends
MODELTRENDS.AI
Model Trends
modeltrends.ai
The wordmark
The mark, then the name in Helvetica: Model in accent blue and bold, Trends in white and light, .ai in grey and light, all on one baseline with no spaces.
Model
Trends
.ai
Model
Trends
.ai
© ModelTrends.ai
|
Made in Japan
|
© 2026