Feyn, a startup founded by Shreyash, has publicly released Pulpie, a set of 'Pareto optimal' models designed to clean web content. The models strip boilerplate elements like ads, footers, and sidebars from raw HTML, returning only the main content in either HTML or Markdown format.
The company claims its approach offers a significant economic advantage. According to Feyn, cleaning 1 billion webpages would cost $7,900 with Pulpie, compared to $159,000 using Dripper, the current leading extractor. This represents a more than 20-fold reduction in cost for large-scale web processing.
Pulpie's efficiency stems from a different architectural approach. While leading extractors are decoders that generate output token-by-token, Pulpie uses an encoder that runs a single forward pass over the full input. This makes it compute-bound rather than memory-bound, allowing it to run optimally on more affordable hardware.
The motivation for the project, as stated by Feyn, came from building a 'deep research harness' that required extensive web data processing. The company has provided a side-by-side comparison tool on Hugging Face for users to test Pulpie against competitors.
The broader impact is most pronounced for AI research and data-heavy web scraping operations. If the cost claims hold true, it could lower barriers for smaller teams working on large-scale web extraction tasks. However, the model's performance on highly dynamic or JavaScript-rendered pages remains unaddressed in the announcement.