kev-ministral · Ministral-3-3B-Base + Kev decision head, in your browser

One download, one WebGPU load, three task shapes: the stock base model (int8) generates text, and the same decoder computes Kev decisions when its built-in LoRA branches are switched on (lora_scale = 1) and a pointer head reads the hidden states. Decisions: Mind2Web unseen websites 69.8% in this browser build (PyTorch 69.0%). Needs Chrome with WebGPU and ~9 GB of free memory.

Example (live pages captured 2026-10-04, and held-out game runs: none seen in training)

 Use my screenshot…

Screenshot (boxed controls are labelled with their refs)

decide without the screenshot

State (what webrunner tells System 1: goal, history, page controls) Question Options — one per line, key: description

Decide

Image

 Use my image…

Prompt (the image goes where [IMG] is; this is a base model, so start the sentence you want finished) [IMG] This screenshot shows a web page with

Max tokens

Temperature

 Generate

Output