OptiQ

OptiQ is an MLX-native toolkit for running large language models on Apple Silicon. No PyTorch, no CUDA, no cloud.

Its main job is mixed-precision quantization. Rather than applying one bit-width to the whole model, it measures each layer's sensitivity and assigns bit-widths per layer, keeping 8-bit where accuracy depends on it and dropping to 4-bit or 2-bit where it does not. On a sparse mixture-of-experts model this matters a lot, since the routed experts are most of the weights.

It also does LoRA fine-tuning and serving. optiq serve exposes an OpenAI and Anthropic compatible endpoint with a mixed-precision KV cache, prompt caching and tool-call healing, and can stream expert weights off SSD so a model far larger than RAM still runs.

OptiQ Code is a terminal coding agent built on the same stack, aimed at local models.