On-device vision

WebBrain VL

Liquid’s LFM2.5-VL-450M, fine-tuned to turn a 1280×720 viewport into WebBrain’s six-section observation: purpose, exact visible text, inputs, state signals, blockers and unknowns. It runs inside the extension through ONNX and WebGPU.

Model on Hugging Face Dataset

44/100strict, deployed ONNX build

0/100 → 36/100base to fine-tune (PyTorch)

50,000browser screenshots in the dataset

0.81 GBpackage that runs in the browser

Why it matters

A cheap text-only planner can’t see a page. WebBrain VL gives it local eyes without sending screenshots anywhere, and it is now WebBrain’s default local vision model.

Training

LoRA rank 16 on a 452,847,616-parameter base, training about 0.91% of the weights, for one epoch on a single RTX 4090. Round 1 used 16,646 screenshots from WebSight, MiniWoB++ and ScreenParse; round 2 grew to exactly 50,000 by adding 33,354 synthetic HTML/CSS screens rendered in a real browser. Teacher labels came mainly from Qwen3.6-35B-A3B.

100 held-out browser screenshots

Model	Strict	Mean score
Qwen 3.8 27B NVFP4 (reference)	74/100	93.2%
Qwen 3.5 4B	55/100	83.4%
WebBrain VL 2 450M (ONNX)	44/100	76.24%
Gemma 4 E4B	41/100	81.3%
Qwen 3.5 2B	39/100	79.1%
LFM2.5-VL-450M (base)	0/100	4.17%

Task-specific parity, not universal superiority: Qwen and Gemma keep higher mean scores. The ONNX build’s gain over PyTorch (44 vs 36) is a runtime effect.

Models

webbrain-vl-2-450M WebBrain VL 2, from LFM2.5-VL-450M 168 downloads · 5 likes
webbrain-vl-2-450M-onnx ONNX build that ships in the extension 387 downloads · 0 likes

Datasets

webbrain-vl-2-450M-dataset 50,000 browser screenshots with six-section observations 587 downloads · 2 likes

Caveats

Teacher labels may contain OCR errors or hallucinations; high-impact visual claims should be corroborated.