Hi Hugging Face team,
I built a program to continuously monitor download counts for popular public models across platforms. On the Hugging Face Hub, it queries the API using downloadsAllTime
(not the 30-day downloads
field) and stores one snapshot per model per day.
Recently I noticed that ** downloadsAllTime can decrease** when queried on consecutive days. Since this field is documented as the cumulative download count since repo creation, I wanted to check whether this is expected — or whether it might indicate a retroactive adjustment on the Hub side.
from huggingface_hub import model_info
info = model_info("Qwen/Qwen3.5-4B", expand=["downloadsAllTime"])
print(info.downloads_all_time)
I always pass expand=["downloadsAllTime"]
, as recommended in this thread.
Below are daily snapshots from my program (same Hub API query each day). These are all relatively new releases (Qwen 3.5 series and DeepSeek V4), yet they already show the same pattern:
1. Qwen/Qwen3.5-4B
| Date | downloadsAllTime |
|---|---|
| 2026-06-08 | 19,704,972 |
| 2026-06-09 | 19,329,213 |
| 2026-06-13 | 18,001,904 |
Current value (queried on 2026-06-13): 18,001,904
Change 2026-06-08 → 2026-06-13: −1,703,068 (−8.6%)
2. Qwen/Qwen3.5-35B-A3B
| Date | downloadsAllTime |
|---|---|
| 2026-06-08 | 11,183,510 |
| 2026-06-09 | 11,018,594 |
| 2026-06-13 | 10,782,371 |
Current value (queried on 2026-06-13): 10,782,371
Change 2026-06-08 → 2026-06-13: −401,139 (−3.6%)
3. deepseek-ai/DeepSeek-V4-Pro
| Date | downloadsAllTime |
|---|---|
| 2026-06-08 | 6,692,783 |
| 2026-06-09 | 6,288,488 |
| 2026-06-13 | 5,999,035 |
Current value (queried on 2026-06-13): 5,999,035
Change 2026-06-08 → 2026-06-13: −693,748 (−10.4%)
All three models show a monotonic decrease across these snapshots. The repos themselves were not deleted or renamed during this period.
The program is meant to track how popular models are growing over time across the ecosystem. For Hugging Face, it computes day-over-day differences of downloadsAllTime
to estimate daily new downloads and plot trends.
Lately, several models show negative daily increments because the cumulative total itself went down between snapshots. That makes the charts hard to interpret — readers naturally expect download counts to only increase.
I understand that the downloads
field (30-day rolling window) can decrease, but I had assumed downloadsAllTime
was non-decreasing unless there was an explicit data correction. I’m trying to figure out whether I should change how my program handles Hugging Face stats, or whether this reflects something on the Hub side.
Is it expected that downloadsAllTime can decrease over time? If so, under what circumstances (deduplication, bot filtering, bug fixes, counting rule changes, etc.)?
Has there been a recent retroactive adjustment on the Hub that could explain broad downward shifts across many public models in early/mid June 2026?
Is there a recommended way for downstream consumers to compute reliable daily download trends from the public API, given that all-time totals may be revised?
Happy to share more model IDs, raw API responses, or a minimal reproduction script if that would help investigate.
Thanks!