Luma v2.9 is a language model based on a transformer architecture with approximately 10 million parameters, designed to be trained and run completely locally. This means it doesn't require a cloud connection or send data via telemetry, giving users full control over their data and the model.

Key Features

  • Custom Training: Luma v2.9 can be trained on specific datasets, organized into three folders: Core, Knowledge, and Conversations. Data weights can be manually defined.
  • Local Execution: The model is designed to run on consumer GPUs or CPUs, without exotic dependencies. It is built with PyTorch.
  • Small Size: Luma v2.9 is intentionally small, with the goal of providing a specialized and focused model, rather than a replacement for larger models like GPT-4 or LLaMA.