Description
The cklxx/laya-browser model is a specialized feature extraction model, boasting 0.3 billion parameters, designed to function as a browser-agent decision head. It is a fine-tuned version of the `laya` model (convaiinnovations/laya), which operates as a non-autoregressive "System 1" decision model. Unlike traditional autoregressive models, `laya` uses a single bidirectional encoder pass to answer several typed questions (choice, score, noul) with calibrated probabilities, without generating text.
Read More
This particular repository fine-tunes `laya` specifically for browser automation tasks, serving as a drop-in replacement for `browser-use/jev-ultrafast`. It processes `/v1/systemone` requests, which mirror `laya`'s `predict(state, questions)` function. In every step, the model performs one forward pass, taking approximately 22 milliseconds on a full GPU clock, to select the appropriate operation (such as CLICK, TYPE_TEXT, SELECT, PRESS_ENTER, SCROLL, or DONE) and its corresponding target element on a webpage.
The model was trained and evaluated locally on an RTX 4070 Ti SUPER (16 GB) without relying on paid APIs. For text input and dropdown value selection, it integrates with a local Qwen3-8B-AWQ model. The current version, v19s, is available at the repository root, with earlier checkpoints accessible in the commit history. Updates are only released when a new version demonstrates clear improvements.
Evaluation results highlight its performance across various suites. Suite C, which involves 27 multi-step tasks on held-out real sites (including search, filters, sort, and open operations), shows v19s achieving approximately 20–26% success, a significant improvement from v17s's 2/54. Suite B, with 18 tasks on 18 held-out sites, maintains a perfect 54/54 score for both v17s and v19s. In Suite A, comprising 16 original real-site tasks, v19s scored 39/48. For webgym held-out synthetic forms (7 kinds × 10), v19s achieved 30/70, doubling v17s's 15/70. The model also demonstrates strong performance in held-out decisions on unseen sites (WebChain) for click operation and target top-1 accuracy.
Latency for a single decision is typically 22–27 ms on an RTX 4070 Ti SUPER at full clock. However, if the GPU drops to idle clocks between decisions (due to waiting for pages to load), subsequent decisions can take 75–270 ms. Users can lock the minimum SM clock to mitigate this. The model is integrated using `transformers` and `torch`, shipping its own code. It can also be loaded via the upstream `laya` package. Key changes since v17s include harness improvements, such as waiting for navigation to complete, an "undo guard" to prevent repetitive actions, better handling of form field data, and scrolling options into view. Training data was also expanded with WebChain human trajectories and DAgger on webgym. Current limitations include stopping too early in multi-step tasks and issues with pages requiring extensive scrolling or sites blocking headless Chromium.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
cklxx/laya-browser Highlights
Non-autoregressive decision model for browser automation
Single bidirectional encoder pass for typed questions
Calibrated probabilities for choice, score, and noul answers
Drop-in replacement for browser-use/jev-ultrafast
Picks browser operations (CLICK, TYPE_TEXT, SELECT, PRESS_ENTER, SCROLL, DONE)
Identifies target elements for operations
Fast decision latency of ~22 ms on full GPU clock
Trained and evaluated locally on RTX 4070 Ti SUPER
Integrates with local Qwen3-8B-AWQ for text input and dropdown values
Supports multi-step tasks on real-world websites
Includes an "undo guard" to prevent repetitive actions
Improved handling of form field placeholders, types, and patterns
Offers options scrolled out of view by scrolling them into view
Trained with WebChain human trajectories and DAgger on webgym
Apache-2.0 license for the model code
Getting Started with cklxx/laya-browser
Access Model: Download the cklxx/laya-browser model from Hugging Face.
Set Up Environment: Ensure `torch` and `transformers` libraries are installed in your Python environment.
Integrate via API: Use the `decide` function to build requests in the model's trained format and run a forward pass.
Utilize as Replacement: Implement `m.systemone(body)` to answer `jev-ultrafast /v1/systemone` requests.
Load with Laya Package: Alternatively, load the model using `laya.load("cklxx/laya-browser")` with the upstream `laya` package.
Apply Harness Improvements: Apply the `code/jev-ultrafast.patch` to `jev-ultrafast` for improved evaluation and training setup.
Optimize Latency: Optionally, lock the GPU's minimum SM clock using `sudo nvidia-smi -lgc 2100,3135` to reduce decision latency.
cklxx/laya-browser's Use Cases
- Automated Browser Navigation
- Web Data Extraction
- Form Filling Automation
- Website Testing
- Agent-Based Web Interaction
- Research Data Collection
- Repetitive Online Tasks






