Mirai Labs

Convert and optimize your model for iPhone, iPad and Mac. One command to get your model running on 2 billion Apple devices.

The uzu Mac app allows users to run recommended models, compare uzu with other engines, and integrate it via the SDK, app, or a coding agent.

25 Qwen, LFM, Muse-Glimmer and d3-flash checkpoints in 4-bit (Mirai-M) and 8-bit (Mirai-L) for Apple Silicon, with measured uzu tok/s for every checkpoint.

The uzu Mac app lets users chat privately and securely with AI models directly on macOS and Apple Silicon devices.

Mirai-M (4-bit) and Mirai-L (8-bit): a co-designed quantization and inference stack with 40-60% more tokens per second than llama.cpp and MLX at the same quality level.

Research and engineering notes from Mirai Labs: Mirai Quantization, sparse buffers for the KV cache, and autoregressive drafting for local LLMs on Apple Silicon.

How uzu backs the KV cache with Metal sparse buffers so memory grows page by page instead of committing 4.5 GiB up front for a 32k-context Qwen3-4B.

Mirai Labs releases speculative decoding for uzu, initially supporting Qwen3.6 27B and planning to add Qwen3.8 27B and Muse Glimmer. On Apple M5-series chips, uzu outperforms MTPLX by nearly 2x and llama.cpp by over 3x at comparable quantization levels, excelling in mathematical reasoning and coding tasks.

Apply for the inference engineer role at Mirai. Join a small, senior team building the full on-device stack to achieve realtime local intelligence.

Apply for the member of technical staff role at Mirai. Join a small, senior team building the full on-device stack to achieve realtime local intelligence.

Mirai Labs develops the full on-device AI stack to enable realtime local intelligence.

Mirai Labs provides contact via email at contact@trymirai.com or a form for inquiries about their work.

Mirai Labs builds uzu, a Rust LLM engine for iPhone, iPad and Mac, and 4-bit/8-bit Qwen, LFM, Muse-Glimmer and d3-flash models benchmarked vs MLX and llama.cpp.

GitHubRepository owner

Mirai Labs introduces **uzu**, an open-source engine that accelerates local LLM inference on Apple Silicon by up to **4x**. The tool uses **tree-based speculative decoding** to process multiple tokens in a single forward pass, achieving **92.1 tokens/sec** for Qwen3.5 9B on a base M5 MacBook Pro. This outperforms competitors like llama.cpp (22.0 tokens/sec) and MLX (25.1 tokens/sec) by optimizing decoding structure rather than relying solely on hardware.

Oct 2026Thaki Cloud Tech Blog | ThakiCloud | 다키클라우드 기술 블로그News article

Mirai Tech Inc. secures $10 million in funding to build infrastructure enabling AI models to operate on smartphones and laptops, eliminating reliance on remote server requests. The company aims to streamline AI processing by running computations locally on user devices. This shift could enhance privacy and reduce latency for applications relying on AI.

Feb 2026dev.uaNews article

Uncork Capital adds **Mirai** to its investment portfolio, revisiting an earlier thesis on edge computing. The firm highlights a delayed but now opportune moment for companies leveraging decentralized, low-latency infrastructure. Mirai’s focus aligns with Uncork’s long-standing bet on distributed systems and real-time processing applications.

Feb 2026MediumNews article