Skip to content
NEURAX
Download

Security & privacy

Local by default, with clear data routes

Model analysis, supported dataset profiling and training run on the selected machine. Network access is used when you choose it: public Hugging Face or ModelScope imports fetch model configuration, optional sign-in checks your plan, and the Agent sends its prompt and project design/data-profile context to the provider you configure. NEURAX does not send raw dataset contents or local paths to that provider; its own terms, logging and charges apply.
The NEURAX studio running on the local machine: GPT-2 Medium on the canvas with its diagnostics computed on-device

  • Local processing

    The desktop engine compiles designs, profiles supported datasets and runs supported training on the selected machine. A Hub import or sign-in is a separate, explicit network request.

  • Import real published architectures

    Enter a public Hugging Face or ModelScope model ID or URL. NEURAX fetches its supported configuration and rebuilds the architecture for inspection and analysis; this import path does not download the checkpoint weights.

  • The Agent uses your configured provider

    When you invoke the optional Agent, its request includes your prompt and project design/data-profile context and is sent to the endpoint you configure. NEURAX does not proxy the request; the provider processes it under its own terms and may log it or charge for usage.

  • No raw dataset file in the Agent request

    The dataset is read locally. The Agent request can include the dataset profile needed for reasoning, but it does not include the raw dataset contents or local file path.

  • Sign-in is optional for local work

    You can use local design and analysis without signing in. If you sign in, the account service checks your identity and plan; that account exchange is separate from local model and dataset processing.

  • The model you approved is the model that trains

    Generated code is built, counted and pushed one batch through before a run starts. Code that disagrees with the design is refused rather than trained.

  • Nothing locked in

    A run produces ordinary PyTorch files and weights in your own project folder. Stop using NEURAX tomorrow and everything you trained keeps working.

Found a vulnerability? Please report it privately to contact@neuraxs.dev.