Executive Summary
Ten daily captures of the HuggingFace trending board hold 301 slots and 64 different models. The average model appears on 4.7 of them and the median appears on four. Nine never left across the whole stretch, and only one of the nine gained attention while it sat there, ccore’s Qwen3.5-4B-3D-Coder, up 53 percent from 271 to 415.
The nine that stayed are not the strong ones. Five of them finished with a lower trending score than they started, and the deepest fall was DeepSeek-V4.1-Flash-1017, down 76 percent from 872 to 205 while its downloads doubled from 131,062 to 277,323. Turnover runs near four names in and four out per day, and this morning brought four first timers, three of them under 1,400 downloads. Tenure says more than rank does here, because rank resets and tenure does not.
Today’s board exchanged four names for four, which is an ordinary morning. speridlabs/idol took rank one on its ninth capture with 20,689 downloads. moonshotai/Kimi-K3.1-Pro held rank one yesterday at a score of 1,017 and answered this morning at 535, a 47 percent drop in a single day. Its downloads rose 22 percent over the same interval.
The two numbers run on different clocks. The score tracks attention, the download counter tracks installs, and across ten days of captures the two almost never moved together. What this trending feed actually measures is how long a repository stays in front of people.
The answer, across 301 slots and 64 models, is 4.7 days on average and four at the median. Only two of the 64 came back after dropping off.
Eleven models appear in one capture and vanish
The bottom of the distribution is where the board does its housekeeping. Eleven of the 64 models appear in exactly one of the ten captures, six of them on the opening day. Eight appear twice. Nine appear three times and nine more appear four. Getting to five appearances still leaves a repository as one of only 27 that went that far, fewer than half of everything seen in the window.
This morning’s four arrivals show what a first capture looks like. The xtools-at repository landed at rank three on 1,369 downloads. NousResearch’s Arete-27B came in at fifteen with 989. PrunaAI’s P-H3-Image closed the board at thirty with 789. None of the three carries an install base. Each is ranked on the first burst of interest a new repository gets and on nothing else.
The exits were louder than the entrances. baidu’s Unlimited-OCR-2 left the board from rank twenty seven carrying 310,955 downloads. microsoft’s Mage-VL-Cowork left from thirty. A repository with a third of a million downloads can lose its place overnight, which is the clearest evidence that the ranking is not reading load.
The nine survivors lost ground while their downloads climbed
Nine models held a slot in all ten captures. Five of them finished with a lower score than they started, and the median of those five fell 42 percent. black-forest-labs’s flux-3-image-klein-comfy went from 352 to 257. NVIDIA’s Nemotron-3.5-Transcribe-30B-A4B went from 406 to 290. mistralai’s Mistral-Large-4-1T-A52B fell from 316 to 184.
DeepSeek-V4.1-Flash-1017 is the cleanest case. It opened the window at rank one with a score of 872 and closed it at nineteen on 205. Its download count went the other way the whole time, from 131,062 to 277,323. Three community rebuilds of the same release sit at fourteen, twenty three and twenty six today, so the family holds four of thirty slots while the parent’s own score lost three quarters of its value.
All nine added downloads across the window. That is the problem with using the counter to judge anything. Download accumulation runs close to constant, so it cannot separate a model people run from a model nobody deletes, and attention cannot separate a launch from a rally.
Tenure plus a rising score is the only combination worth tracking
The one survivor that gained attention deserves the look. ccore’s Qwen3.5-4B-3D-Coder went from 271 to 415 on the score, a 53 percent rise, while its downloads went from 12,237 to 221,506. It has been on the board for ten captures and it is still climbing.
That combination is rare here. Ten days produced one clear case of it. A repository that holds a slot for six captures and adds installs on each is something teams are building on. A repository that spikes to rank one and halves by morning is an announcement. A screenshot of the board cannot tell them apart, and the difference is the whole cost of a wrong bet on a GPU fleet.
Three questions to ask before taking weights off this chart. Does the score hold once launch traffic leaves? Do downloads keep compounding past the second week? Has anyone outside the original account shipped a build, because a model with only a parent repository is still a demo?
Our earlier reading of how the board weights momentum over load covers the measurement itself. The ten day view adds the part that matters for a schedule, which is how fast a spike decays and how rarely it comes back.
Get the next one before it is old news
Independent analysis of cloud-native infrastructure, Kubernetes and data center economics. No vendor spin.