lmxlab

Overview

lmxlab gives GPT, LLaMA, DeepSeek, Gemma, Qwen, Mixtral, and related model families a consistent interface in Apple MLX. The shared components make it easier to compare architecture choices without maintaining a separate training stack for every model.

I use it for local experiments on Apple Silicon, including model architecture work, LoRA fine-tuning, and preference optimization. Unified memory allows many of those experiments to run without a remote GPU.

Technical highlights

• Reusable transformer and mixture-of-experts architecture components
• MLX-native execution for Apple Silicon
• LoRA, DPO, GRPO, and related training workflows
• Command-line model inspection and parameter counting
• Published Python package and API documentation

Topics