WebBrain Research

Small models that see, decide and act in your browser.

A browser agent shouldn’t need a datacenter. We build small, specialized models that run where the page is, shrink strong ones until they fit, and test every model we can find on the same frozen browser tasks.

Everything here is open: the weights and datasets on Hugging Face, the OpenBonsai pipeline, and the benchmark results, failures included.

32 open models and datasets5.2k downloads143 likeson Hugging Face, October 30, 2026

Compression method

OpenBonsai

Open-source ternary quantization: any BF16 model or fine-tune becomes weights of only −1, 0 and +1. Qwen3.5-9B shrinks from ~18 GB to ~2 GB.

~18 → ~2 GBQwen3.5-9B, BF16 to ternary~5 hto compress on one RTX 5090 Read the introduction On-device model

Compass VLD

One small model that sees the screenshot, reads the request and decides: vision, language and decision together, so the browser needs no separate captioner, planner and judge.

120/126page-state decisions, zero false completions0.94 sper decision, in the browser Read the introduction Tool-routing models

WebBrain Compass

Small fine-tunes that pick the next browser tool and its arguments, so routine steps don’t need a large planner.

9 → 16strict routing matches /89, Tiny v2 vs its base Read the introduction On-device vision

WebBrain VL

A 450M vision model that turns a screenshot into what a planner needs: text, inputs, state, blockers. Screenshots stay on your computer.

0.81 GBin-browser package Read the introduction On-device model

d1 decisions

A model that doesn’t write text at all: it answers yes/no and multiple-choice questions about the page, like “is the cart ready?”, in one pass on WebGPU.

144/144decisions agree, FP16 vs FP32 Read the introduction Evaluation

Open benchmarks

Frozen, public tests of which models can actually drive a browser: 100 planner prompts, 100 vision screenshots, compact routing, all with failures published.

100prompts, frozen since May 23, 2026 Read the introduction Model surgery

Giving open models eyes

Strong text-only models can’t see a page. We attach a frozen vision tower with a small trained projector: 30.68M new parameters give a 33B model sight.

30.68Mparameters trained to add vision to Laguna XS 2.1$0.073per replay, against $0.670 for Qwen 3.6 27B Read the introduction Quantization

Fast local serving

Which formats and engines make strong models fast enough for an agent loop on your own GPU: NVFP4, EXL3, ONNX.

10.2 → 1.76 smedian step, Qwen 3.6 27B in NVFP4 Read the introduction Tiny classifier

SafeSocial

A 13.6 MB image classifier that runs on the device and blurs social-comparison triggers on Instagram, without waking an LLM for every image.

13.6 MBEfficientNet-Lite0, runs locally Read the introduction Agent method

Vision mode

Driving the browser from screenshots alone, clicking by coordinates like a person, for pages whose structure is missing or misleading.

Autodefault for Act runs with image-capable models Read the introduction

Three ideas run through all of it

Specialize, then shrink
A 2B model trained for one job can match a model ten times its size at that job. Compass Tiny v2 sits one strict match behind Qwen3.8-27B at routing; WebBrain VL 2 beats Gemma 4 E4B on strict screenshot reading at 450M.
Run where the page is
Vision, decisions and routing run inside the browser through WebGPU and ONNX, so screenshots and page content stay on your computer and the agent keeps working offline.
Measure in public
Every model faces the same frozen prompts and screenshots, and we publish what didn’t work alongside what did, so the next model is comparable with the last.