Felix Rieseberg Releases Free macOS App to Train GPT-2-Scale Models Locally
Felix Rieseberg, a developer at Anthropic, has released "Language Model Builder," a free macOS application designed to help users train GPT-2-scale language models from scratch on local hardware. The tool allows users to train small-scale models without needing complex cloud setups. This application lowers the barrier to entry for developers and AI enthusiasts who want to understand the mechanics of LLM training firsthand. By enabling local training on consumer hardware, it democratizes access to hands-on machine learning education. Using default settings, users can train a model that writes coherent text in about a day, while a GPT-2-small-class model with 100–150 million parameters takes about a week on a MacBook Pro M5 Max. The creator notes that the tool is intended for educational purposes and cannot be used to train massive frontier models.
## BACKGROUND
GPT-2 is a transformer-based language model originally released by OpenAI in 2019, with its "small" variant containing approximately 124 million parameters. Training such models typically requires understanding complex machine learning frameworks and managing GPU resources. Local training on macOS has become more feasible recently due to Apple's unified memory architecture and specialized hardware accelerators.