Mirai

Mirai is an on device AI lab for Apple Silicon. The company builds the full stack needed to run language models locally on iPhone, iPad and Mac: models designed for memory bound execution, a quantization scheme co designed with those architectures, and an inference engine written in Rust called uzu. The premise is that on device inference is a different problem from cloud inference. With a single request processed at a time, memory bandwidth rather than raw compute is the bottleneck, so Mirai tunes models and runtime together to minimize memory footprint and make full use of the neural accelerators. The stated goal is very high token throughput, around 1,000 tokens per second, so that AI features feel instant and can run continuously inside an interface instead of behind a chat box. The company publishes 4 bit and 8 bit versions of Qwen, LFM and its own Muse Glimmer models and benchmarks them against MLX and llama.cpp. It is built for developers and product teams who want private, low latency inference on Apple devices, and the uzu engine is available on GitHub. The site also references a Mac app and an SDK. No prices are published. The pricing page is a contact form, and commercial terms are discussed directly with the team.