Bought charts? We investigated who's liking Hugging Face's top trending models

What this is about

If you go looking for open AI models on huggingface.co, the first thing you see is the "Trending" list. Whatever sits at the top ends up in newsletters, automated news digests and Reddit threads, and gets downloaded. That makes the list a kind of music chart for open AI. Over the weekend, the community on Hugging Face and Reddit raised an uncomfortable question: how many of these likes are real? The trigger was discussion #97 on the video model LTX-2.5 from Lightricks. At the same time, a conspicuous model is sitting at #1.

We analyzed Hugging Face's public data ourselves. A note up front: we didn't count complete lists of thousands of likers. We took samples, mainly the 100 most recent likers of seven models plus excerpts from the middle of the lists. We pulled the data on November 15, 2026. Automated analysis may have introduced the occasional reading error.

How the Trending list works (as far as we know)

Hugging Face doesn't publish a formula for the trending score. A question about it on the official forum has gone unanswered by staff since November 2024. The dataset severo/trending-repos describes Trending as a reflection of "the likes Hugging Face users gave in the last week." Our snapshot is consistent with that. For brand-new models, the score almost exactly matches the total number of likes: Haru-4B has 283 likes and a score of 267, Edge0-145B has 278 likes and 250. Downloads apparently barely matter. The Comfy-Org repo of MiniMax-H3-Image was downloaded almost two million times in 30 days and still only scores 366 points.

Our take: at its core, the score is a counter of fresh likes, possibly time-weighted. If you can manufacture likes, you can manufacture the charts.

The snapshot: the top 20 on November 15

The ratio of likes to downloads isn't proof. Models that are only three days old naturally have few downloads, and hardly anyone downloads huge models like Kimi for personal use. Still, the top model stands out. It's 16 days old and has only 9 billion parameters, so it runs on any decent gaming GPU. It gets 276 likes per 1,000 downloads, roughly 90 times as many as Qwen3.8-27B. According to its Model Card, it's a version of Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B "uncensored" with the Heretic tool. Heretic is a method that strips refusals out of a model: instead of refusing 100 out of 100 test prompts, it refuses only 3. According to the API, the base model itself has a grand total of 82 likes and 2,515 downloads.

What the liker lists reveal

The /api/models/{repo}/likers endpoint is public. For each liker, it returns an _id, a so-called MongoDB ObjectID. Its first 8 hex characters are a Unix timestamp, revealing when the account was created. For example, 6af90000 corresponds to November 15, 2026, 00:42 UTC. So anything starting with 6af8 or 6af9 was at most about 36 hours old when we pulled the data.

The result contradicts the obvious assumption. "Lots of likers created their accounts the same day" applies to the suspicious top model, but just as much to Qwen3.8-27B and to Meta's year-old segmentation model sam3. On its own, this metric doesn't separate real from fake.

Two individual cases show why. The account BlackDevil38 was created on November 15 at 06:13 UTC. Three minutes later, within 52 seconds, it liked exactly the #1, #3 and #4 models on the Trending list and nothing else. The account Ppeerapon, by contrast, has existed since July and has 106 likes spread over months, mostly embedding and translation models. On November 15, it liked eleven trending models within eleven minutes, in almost exactly the order of the list. That's how someone behaves when they click through the charts from top to bottom. Our take: some of the fresh likes are genuine newcomer behavior, and the Trending list reinforces itself. Once a model is at the top, it gets more likes, including real ones.

The clearer traces: assembly-line accounts

The series that look like mass production are more clear-cut. Roughly in the middle of the top model's liker list, there are around 30 consecutive accounts with names like jimmyhess1993, lori1992, bryan0427, kristen0718 or robertdixon1998. Their timestamps all fall between 19:17 and 19:36 UTC on November 6. The account jimmyhess1993 has given exactly three likes in total: one on November 8 and two on November 12, for the top model and PrunaAI/P-H3-Image, five seconds apart.

The same pattern appears in other repos: on sam3 (adamcole1999, jacob1993, rhondahaas0418), on speridlabs/idol, on a DavidAU model and on MiniMax-H3. There, nine accounts following the pattern "first name + last name + 6 hex characters" (such as lunahoward19f6ee) are liking the model, all created on November 12 between 04:05 and 05:38 UTC. Before that, lunahoward19f6ee had liked a Whisper model, a dataset and a Space. Our take: this fits bot accounts that camouflage themselves with scattered likes. The patterns resemble the ones HiddenLayer described in May ("firstname-lastname###").

One important point: if the same account factory likes many popular repos, its presence doesn't tell you who paid for it. That goes for the top model, too. We have no evidence that xtools-at was involved.

The LTX-2.5 case

For LTX-2.5, user EclipseMist analyzed the complete lists in discussion #97:

For H3, this looks like a normal launch: lots of newcomers in the first month, fewer afterward. For LTX-2.5, new-account likes rose for three months straight, while downloads stagnated at around 2.6 million a month, according to the Reddit post. Public digests on GitHub confirm the rise: 8,377 likes on October 19, 10,387 on November 3, 12,014 on the day we pulled the data. Our sample from the middle of the list shows another wave: 34 of 60 consecutive likers have accounts created on October 4 between roughly 13:20 and 21:20 UTC. The names, however, look individual rather than templated.

Nobody in the discussion mentions one confounding factor. According to the API, LTX-2.5 is "gated": to download it, you have to log in and accept a license. MiniMax-H3 isn't gated. So some of the new sign-ups could be real users creating an account just to download it. That doesn't explain months of growth while downloads stayed flat, though. We didn't verify the benchmark claims in the Reddit post (LTX-2.5 at 72 percent versus 87 percent for H3 in a test by IFM). Lightricks and Hugging Face still haven't commented in the thread.

Why this can get dangerous

The Open-OSS/privacy-filter case in May showed how risky this is. The repo posed as OpenAI's "Privacy Filter" and, according to HiddenLayer, reached #1 in under 18 hours, with around 244,000 displayed downloads and 667 likes. 657 of those likes came from accounts with templated names. The loader.py script installed an infostealer on Windows, malware that grabs passwords, session cookies and crypto wallets. Hugging Face deleted the repo. According to Decrypt, the platform hasn't said whether it will vet trending repos more closely in the future. The analysis by Harper Foley sums it up:

A popularity ranking is not an identity.

Hugging Face's Content Policy explicitly prohibits manipulating Hub metrics ("e.g., rewards in exchange for likes"), as well as "excessive bulk activity" and unauthorized bot access. Whether Hugging Face filters out likes from young accounts when computing the score isn't documented. Our samples suggest it doesn't. As for the current top model, there's no sign of malicious code: the Model Card lists only standard files in Safetensors format and no custom loading scripts.

The bigger picture

Our take: Hugging Face's Trending list is cheap to game. It essentially counts fresh likes, and accounts can apparently be created by the dozen within minutes. The data shows templated accounts in at least five of the repos we checked. But it also shows that the often-cited metric "share of young accounts" says little on its own. For Qwen3.8-27B and sam3, it's about as high as for the top model. Three other signals are more reliable: series with identical naming patterns and timestamps minutes apart, a like-to-download ratio far outside the norm, and like curves that keep climbing months after launch.

Who's behind it remains an open question: the model makers, third parties, or like-selling services that also like other people's models as camouflage. For you as a user, this means "Trending" is no seal of quality. Look at the publisher, the downloads and independent tests. Don't run scripts from unknown repos. Checking the public liker list takes only seconds. Hugging Face could defuse the problem by publishing the formula and giving likes from very young accounts less weight in the score. The platform hasn't commented on this so far.