We just open-sourced NeuralAction, our 4B GUI grounding AI model powering NeuralAgent
We just open-sourced NeuralAction, our 4B GUI grounding model powering NeuralAgent.
The goal is faster, lower-cost computer use: keep planning and application knowledge with the frontier model, and delegate repeated visual grounding to a small specialist.
The supervisor handles planning and application knowledge. NeuralAction locates the next visual target. Checkpoints verify progress and return control to the supervisor when the workflow needs adjustment.
We’re releasing:
- Open model weights under Apache 2.0.
- A Python desktop demo that executes actions while the supervisor streams its plan.
- Independent provider configuration for supervisor and checkpoint models, supporting OpenAI, Anthropic, Gemini, and OpenAI-compatible endpoints.
NeuralAction reports 91.7% on ScreenSpot-v2 and 49.9% on ScreenSpot-Pro. In a separate grounding latency measaurement, it achieved 258 ms p50 on an A100 with 1280px input over 30 warmed-up requests.
The attached demo shows events in real time.
Download the model:
https://huggingface.co/withneural/NeuralAction
Code and documentation:
https://github.com/withneural/NeuralAction
Experience it through NeuralAgent or Aura by Neural:
NeuralAgent:
https://getneuralagent.com
Aura by Neural:
https://aura.withneural.com
Replies