Web scraping just got a major cost breakthrough. A startup called Feyn has released Pulpie, a family of open-source models that cleans raw HTML by removing boilerplate like ads, navigation bars, and sidebars to expose only the main content. Unlike traditional decoders that generate output token by token—each requiring a full model read—the Pulpie models use an encoder design, processing the entire HTML in a single forward pass to label each block.
The efficiency gains are dramatic. Cleaning one billion web pages costs just $7,900 with Pulpie, compared to $159,000 with Dripper, currently the leading extractor. This price advantage stems from Pulpie’s compute-bound operation versus the memory-bound nature of decoder models. Cheaper GPUs often have more compute capacity than memory bandwidth, making Pulpie easier to run optimally.
Why cleaner web data matters
Feyn’s founders built Pulpie after a costly oversight. While developing a deep research tool, their search API accidentally returned a Google Pixel ad in the results. The ad infiltrated an LLM’s context window and ended up in a user-facing answer, exposing the risks of noisy data in AI pipelines. That embarrassing moment highlighted a critical gap: high-quality data extraction was either expensive or unreliable.
Pulpie addresses this by offering state-of-the-art (SOTA) extraction quality at a fraction of the cost. The models support output formats like HTML and Markdown, making them versatile for developers integrating clean data into workflows.
How Pulpie outperforms the competition
Traditional content extraction tools rely on decoder architectures, which generate output sequentially. Each token produced requires reading the entire model from memory, a process that becomes increasingly expensive as output length grows. Pulpie flips this model by using an encoder approach:
- Single-pass processing: The model analyzes the full input HTML once, labeling each block as boilerplate or content.
- Reduced memory overhead: Avoids repeated memory reads for token generation.
- Cost efficiency: Runs optimally on cheaper GPUs with higher compute-to-memory ratios.
Feyn compares Pulpie’s performance against Dripper in a side-by-side demo, available on YouTube and Hugging Face Spaces. Users can test the difference themselves, evaluating how each tool handles real-world web pages.
Open-source accessibility and next steps
Pulpie’s models are fully open source and hosted on Hugging Face. Feyn provides detailed documentation on training methodologies and implementation guides to help developers get started. The company’s blog post outlines the full process, from model design to deployment.
For teams struggling with noisy data in AI pipelines or web scraping workflows, Pulpie offers a compelling alternative to existing extractors. Its balance of cost, performance, and accessibility could reshape how developers clean web content at scale.
The future of data extraction may hinge on efficiency as much as accuracy. With Pulpie, Feyn is proving that cleaner, cheaper, and faster doesn’t have to be mutually exclusive.
AI summary
Pulpie, web sayfalarından reklam ve gereksiz unsurları otomatik olarak temizleyen açık kaynaklı bir model ailesi. Maliyetleri yüzde 95 düşürürken kaliteyi koruyor. Nasıl çalıştığını ve kullanımını öğrenin.
