Mirai

Mirai is a high-performance on-device inference engine for Apple Silicon that helps AI developers and model companies convert, optimize, distribute, and run models across iPhone, iPad, and Mac.

Key capabilities include:

• Model conversion and optimization from Hugging Face, including quantization and an end-to-end pipeline to prepare models for edge deployment.
• Offline local inference for text tasks such as summarization/extraction, classification, routing, and translation, enabling private experiences with no internet requirement.
• Hybrid device-cloud routing that runs compact models locally and routes larger workloads to cloud infrastructure when needed.
• Performance-focused runtime features, including speculative decoding (up to 2x faster generation) and reported speedups such as up to 38% faster prompt processing versus MLX, without changing model weights.
• Mirai for Mac, a native macOS chat app for running local models on Apple Silicon, pitched as a faster, simpler alternative to Ollama and LM Studio; Android inference is listed as coming soon.

Pricing: the uzu inference engine is open source under the MIT license and free to use; contact Mirai for commercial partnerships and custom deployments.

What makes it notable: it targets Apple Silicon-first deployment with a single conversion pipeline and SDK approach across Apple devices, plus a Rust-based engine with bindings for Rust, Swift, and TypeScript (via the open-source uzu library, about 3.2K GitHub stars). The company reportedly raised a $10M seed round (February 2026, led by Uncork Capital) and was founded by ex-Reface and Prisma team members.