
Introducing Folate: Fast, Local DOM Grounding on CPU
Folate is now available—an open-weight, 110M-parameter model designed to map natural-language instructions to web page elements locally on a standard CPU in 10–25ms.
Browser automation agents typically rely on sending full-page screenshots or massive page trees to large multimodal models for every single click. That approach introduces multi-second delays, recurring API costs, and heavy hardware requirements. Folate solves this bottleneck by separating high-level task planning from element selection—allowing an agent's planner to decide what to do while Folate handles where to click in-process.
How It Works#
Instead of generating text or screen coordinates from scratch, Folate treats element selection as a ranking problem. Given an instruction and a list of interactive elements on a page, the model scores each candidate to identify the best match.
- Compact Architecture: Built on a 110M-parameter ELECTRA cross-encoder backbone optimized for web page structure and text alignment.
- Lightweight Runtime: Quantized to INT8 ONNX so it runs natively on standard CPUs without a GPU or external inference server.
- Minimal Footprint: Uses less than 250 MB of RAM, making it practical to run across standard cloud workers, laptops, and local environments.
Benchmarks & Performance#
On the Mind2Web benchmark, Folate v1 (INT8) reaches ~91% Top-1 element grounding accuracy (~86% on unseen websites), outperforming GPT-6 Luna's ~81.8% step-level accuracy by roughly 9–10 percentage points—nearly cutting the error rate in half while running completely offline on CPU:
| Metric | Folate (110M INT8) | GPT-6 Luna |
|---|---|---|
| Deployment & Hardware | Standard CPU (Local, In-Process) | Cloud API |
| Memory Usage | < 250 MB RAM | API-bound |
| Scoring Speed | 10–25 ms | Network + API inference latency |
| Mind2Web Accuracy | ~91% overall (~86% unseen websites) | ~81.8% (Multimodal step selection) |
Including live page extraction and element filtering, a complete step finishes in 500 ms to 2.3 seconds depending on page size, with zero per-token API fees.
Integration & Availability#
Folate is packaged as a standalone library that drops into any Playwright or Puppeteer workflow.
- Live Demo: Test real-time element ranking directly in the browser on Hugging Face Spaces: huggingface.co/spaces/figranium/folate-demo
- Model Weights: Download the ONNX and PyTorch weights on Hugging Face: huggingface.co/figranium/folate
- Source Code & Docs: View setup instructions and benchmarks on GitHub: github.com/figranium/folate
Current Scope#
Folate relies on standard HTML structure, visible text, and accessibility attributes. As a result, it is not suited for <canvas>-rendered interfaces, unlabelled custom icons with no text metadata, or restricted cross-origin iframes. Because Folate handles single-step element selection, multi-step planning and error recovery remain managed by the surrounding automation framework.