senjai/grok-1

Grok-1 is xAI's large language model, the same AI that powers the Grok chatbot, released as open-source code and downloadable model files. Anyone can now download the model's "brain" (its trained weights) and run it on their own hardware, rather than relying on a hosted API. At a practical level, the repository gives you a small script that loads the model and generates text from a sample prompt. You download the model weights via a torrent client, drop them into a folder, install dependencies, and run the script. The code is written in JAX, a machine learning framework, but you don't need to deeply understand JAX to use it, you just need to follow the setup steps. The catch is hardware. Grok-1 has 314 billion parameters, which is enormous, you'll need a machine with substantial GPU memory to actually run it. A typical laptop or consumer desktop won't come close. This is really aimed at researchers, AI teams at companies, or hobbyists with access to serious compute infrastructure (think multi-GPU servers or cloud instances with high-end GPUs). One honest tradeoff worth noting: the README states upfront that the "mixture of experts" layer, the part of the model that routes work across sub-networks, is implemented inefficiently on purpose. They chose a simpler approach to avoid requiring custom low-level code, which makes it easier to verify the model works correctly but means it won't run as fast as a fully optimized version could. The project is licensed under Apache 2.0, which means you can use, modify, and distribute both the code and the model weights fairly freely, including for commercial purposes.