I keep a small rule in an image downloader extension that turns Wikipedia thumbnails into the original Commons files. This week the thumbnails on English Wikipedia started coming from a different host, so I went back with curl and measured what changed. Everything below was checked on 14 November 2026 against one file, the Einstein portrait in the English article (Albert_Einstein_Head_cleaned.jpg).
The new thumbnail address
The article HTML now points at thumb.wikimedia.org, with tracking parameters on the end:
//thumb.wikimedia.org/wikipedia/commons/thumb/2/28/Albert_Einstein_Head_cleaned.jpg/250px-Albert_Einstein_Head_cleaned.jpg?utm_source=en.wikipedia.org&utm_campaign=parser&utm_content=thumbnail
The path after the host is the same one upload.wikimedia.org has always used: thumb/, the two hash directories, the file name, then <width>px-<file name>.
The old trick for the original still works, as long as you also change the host. Drop thumb/, drop the last segment and the query, and request it from upload.wikimedia.org:
https://upload.wikimedia.org/wikipedia/commons/2/28/Albert_Einstein_Head_cleaned.jpg
→ 200, image/jpeg, 8,907,528 bytes (4,530 × 5,607)
The same original path on the thumb host does not serve the file:
https://thumb.wikimedia.org/wikipedia/commons/2/28/Albert_Einstein_Head_cleaned.jpg
→ 301 to https://commons.wikimedia.org/
So a rule that only strips thumb/ and the width segment, and keeps whatever host it found, now sends you to the Commons home page with a 301. Mine already rewrote the host, which is the only reason it kept working.
Only some widths exist
Changing 250px to any number used to give you a thumbnail at that width. Now it doesn't. I requested the same file at 14 widths on thumb.wikimedia.org, one second apart:
| Width | Status |
|---|---|
| 120, 250, 330, 500, 960, 1280, 1920, 3840 | 200 |
| 180, 320, 640, 800, 1024, 2560 | 400 |
The 400 is an HTML error page whose message reads "Use thumbnail sizes listed on https://w.wiki/GHai". That short link goes to Wikimedia's list of standard thumbnail sizes. An extra pass earlier the same day on 20, 40 and 60 returned images, and 777 and 1200 returned 400.
If you need an arbitrary width, Special:FilePath still accepts one and rounds it up:
https://commons.wikimedia.org/wiki/Special:FilePath/Albert_Einstein_Head_cleaned.jpg?width=777
→ 302 → Special:Redirect/file/...&width=777
→ final URL .../thumb/.../960px-Albert_Einstein_Head_cleaned.jpg
Without ?width it ends at the original on upload.wikimedia.org, 8,907,528 bytes again. Both final URLs pick up utm_ parameters on the way.
WebP, but only if you ask exactly like Chrome
Same 500px thumbnail, four Accept headers:
| Accept header | Content-Type | Bytes |
|---|---|---|
| none (curl default) | image/jpeg | 81,486 |
image/jpeg |
image/jpeg | 81,486 |
image/avif,*/* |
image/jpeg | 81,486 |
image/webp,*/* |
image/webp | 73,684 |
Chrome's real image header (image/avif,image/webp,image/apng,image/*,*/*;q=0.8) also got WebP. Both files decode to 500 × 619. AVIF in the header without WebP gets JPEG, so the switch looks for WebP specifically.
The WebP response I got had no Vary header, and it was a cache hit (x-cache-status: hit-front, age around 12 hours). I can't see how the caches keep the two variants apart from the outside. What matters for a downloader is the practical result: a browser extension that fetches a thumbnail with the browser's own headers gets a .webp file even though the URL ends in .jpg. If you save it under the URL's name, you get a WebP with a .jpg extension. Sniff the bytes or read Content-Type before you name the file.
The original on upload.wikimedia.org came back as JPEG for me in both cases. Only the thumbnails changed format.
Rate limits
My first loop over the widths had no delay and went to upload.wikimedia.org. By the fifth request the widths on the list were answered with 429, while the widths off the list still got their 400, so that check runs before the limit. With a one-second gap on thumb.wikimedia.org I got no 429s. If you script this, slow down and back off on 429; Wikimedia's user-agent policy also asks for a descriptive User-Agent with contact details.
What I changed
Nothing in the extension: it already moved the host to upload.wikimedia.org and fetched the original. I did rewrite the "edit the URL" steps in the Wikipedia image download guide, since the method that told people to type any width now gives a 400 for most numbers they would try.
One file is a small sample. The 400 page points at a published list rather than a per-file rule, so I expect the widths to hold elsewhere. The WebP switch and the missing Vary header are worth re-checking on your own files before you rely on them.
Top comments (0)