Skip to content
NEURAX
Download

What NEURAX does not do yet

The limits of this version, so none of them surprises you.

NEURAX is in active development: the team publishes new versions regularly, and the limits below are the ones being worked on. Each is stated here so you can plan around it rather than find it in the middle of a project.

  • Training runs on this computer only. A cloud machine (AWS, Google Cloud, Azure) is priced in the analysis, but nothing is trained on it, and Decision does not approve a run there.
  • Data on Amazon S3, Google Cloud or Azure is read for the analysis, not trained on. It is read with your own credentials, a public bucket too (aws configure, gcloud auth application-default login, az login); its profile and figures are real. To train, copy it to this computer and choose the local copy.
  • The installers are not signed yet. Windows may say the publisher is unknown (More info, then Run anyway); macOS asks to allow NEURAX once in System Settings, Privacy & Security (Open Anyway). Check the download against the SHA256SUMS.txt file published with each release.
  • On a Mac, training uses the processor. The Apple GPU is not used; on Windows and Linux an NVIDIA card is used through CUDA.
  • One run at a time on a machine. A second start is refused with the name of the run in progress: wait for it, or stop it.
  • A published model is not fine-tuned from its published weights in the studio yet. Importing a model brings its architecture; a text model then learns its tokenizer from your data and trains from the start.
  • The validation share cannot be set in the studio. It is the one the dataset profile proposes.
  • Some models are drawn rather than imported. A published diffusion pipeline imports its UNet only, graph networks are drawn, and a multimodal model imports its language part only.
  • Memory, time, cost and carbon are estimates. Parameters and FLOPs are exact for every block the compiler prices; a finished run shows its measured figures beside the predicted ones.
  • When a run fails, Training says so without the reason. Ask Neurax AI, which reads it, or open the log in the run's folder.
  • Neurax AI needs your own provider key, and its web research your own Tavily key.
  • The EU AI Act check does not apply the one-third rule yet. Training compute is compared with the systemic-risk threshold (10²⁵ FLOP); the rule for a modified published model is not computed.
  • Docker runs the analysis only. Reading data and training need the desktop application.

Page 2 of 33. This guide is in English for now. It is the same guide that is built into the NEURAX desktop app (the ? button, or F1). Found a mistake? Tell us.