Two Quarters
AI model · Two Quarters

Nandini

Our own AI model, built in the open.

V1

LIVE · BUILT ON GEMMA

Built on Google Gemma. Available inside BrahmAI.

V2

IN DEVELOPMENT

Being built from scratch. That means our own architecture, our own tokenizer and pretraining on data we assemble ourselves, not fine-tuning someone else’s model.

V2 progress

DONE IN PROGRESS UPCOMING
  1. 01DataDONE
  2. 02Architecture and tokenizerDONE
  3. 03Training infrastructureIN PROGRESS
  4. 04PretrainingUPCOMING
  5. 05EvaluationUPCOMING
  6. 06ReleaseUPCOMING

Updates

Last updated:

  1. Training infrastructure work begins

    We are setting up the compute and data pipeline needed to train V2, starting with small runs to check that everything works end to end.

  2. Architecture and tokenizer chosen

    The V2 architecture and tokenizer have been chosen and documented. Next: building the infrastructure to train it.

  3. Training data assembled

    Sourcing, cleaning and filtering of the V2 dataset is finished. Next: settling the architecture and training the tokenizer.