July 6, 2026
Web scrubber or hype washer?
Show HN: Pulpie – Models for Cleaning the Web
This web-cleaning tool dropped a huge price bomb — and the comments got messy
TLDR: Pulpie says it can strip junk from web pages nearly as well as a leading tool while being dramatically cheaper, which matters because cleaner web data can improve AI results. Commenters loved the savings, but also nitpicked the charts, roasted the demo, and questioned whether a simpler free method could do enough.
A new open-source tool called Pulpie just arrived claiming it can clean junk off web pages for a tiny fraction of the price of a top rival, and the comment section instantly split into applause, nitpicks, and classic internet side-eye. The big flex is simple enough for non-experts: when companies scrape the web for AI training or search results, a lot of what they collect is trash like menus, ads, and footers. Pulpie says it can keep the real article text while costing about $7,900 instead of $159,000 to process a billion pages. That got people’s attention fast.
Some commenters were fully in their “chef’s kiss, ship it” era, praising the speed jump and calling the design choice “brilliant.” But this being the internet, celebration lasted about three seconds before the thread turned deliciously picky. One user zeroed in on a chart and basically demanded to know why the graph didn’t start at zero, a wonderfully nerdy form of scandal. Another went after the demo itself, complaining that it looked rough in Mozilla dark mode — because no launch is complete without someone testing the vibes first.
And then came the spiciest jab: why not just convert HTML to markdown and strip ads for free? That comment landed like a “do we really need all this?” grenade. So while Pulpie’s creators were talking about cleaning the web, the community was busy doing what it does best: cleaning the hype, one snarky comment at a time.
Key Points
- •Feyn Labs introduced Pulpie, an open-source family of models for extracting main content from HTML pages by classifying blocks as content or boilerplate.
- •The article reports that `pulpie-orange-small` scores 0.862 ROUGE-5 F1 on WebMainBench, close to Dripper’s 0.864, while using 210M parameters versus 600M.
- •On an NVIDIA L4 GPU, the article says `pulpie-orange-small` processes 13.7 pages per second compared with 0.68 pages per second for Dripper.
- •Using an L4 instance priced at $0.39 per hour, the article estimates a cost of $7,900 to clean 1 billion pages with Pulpie versus $159,000 with Dripper.
- •The article argues that cleaner extraction improves both pre-training and inference, citing AICC results and comparisons showing model-based extraction preserves code blocks and formulas better than heuristic methods.