Research
Our research
We share open proof-of-concepts where the work can help the wider community. Each is a compact model that demonstrates the perception and grounding skills behind Brend. The first is brend-2b, an open-weights GUI-grounding model.
Proof of concept · open weights
brend-2b
brend-2b-260602
An open-weights GUI-grounding model: a screenshot and an instruction in, a click out. A compact GRPO fine-tune of Qwen3-VL-2B, released so the community can run it, fine-tune it, and build on it.
Modelbrend-2b-260602
Base modelQwen3-VL-2B
Parameters2B · BF16
MethodGRPO fine-tune
TaskGUI grounding
LicenseApache-2.0
WeightsOpen · Hugging Face
Production is separate. Enterprise models are trained, compiled, and deployed strictly inside your own isolated, air-gapped environment, and never shipped from ours.
Principles
How we work.
ourselves
- Observation over instruction, learn from watching, not from explicit programming
- Understanding over mimicry, grasp intent and context, not just surface actions
- Build it ourselves, we train our own models and run our own infrastructure
- Open what we can, share work that helps the field move forward
Curious about the work?
We're hiring researchers.