Mirai: Run local models fast on Apple devices
Mirai is an on-device AI platform from Mirai Labs that combines the uzu Rust LLM engine, quantized models and an application layer for Apple Silicon devices.
What is Mirai?
Mirai Labs works on the full on-device stack for Apple Silicon: local model architectures, an inference engine called uzu written in Rust, a quantization scheme, and an application layer that starts with Apple device automation. The company argues that on-device inference is a different problem from cloud serving, since each request runs at batch size 1 with tight memory and bandwidth limits, and it co-designs models and runtime around that constraint. The published material covers uzu for iPhone, iPad and Mac, plus 4-bit and 8-bit Qwen, LFM and Muse-Glimmer models benchmarked against MLX and llama.cpp. The stated goal is realtime local intelligence, with 1,000 tokens per second framed as the point where models become interfaces. Pricing and plan details are not published on the site.
Focus areas: On Device.
Who is Mirai for?
- Developers building AI features for iPhone, iPad or Mac
- Teams evaluating local inference against MLX or llama.cpp
- Product groups exploring device automation on Apple hardware
Is Mirai worth it?
A focused on-device AI stack for Apple Silicon, though pricing and access details still have to be requested
Where Mirai is strong
- Covers models, runtime and quantization as one stack
- Engine and models are aimed specifically at Apple Silicon
- Published benchmarks compare models against MLX and llama.cpp
Where it falls short
- No pricing or plan information is published
- No public documentation, SDK details or platform requirements are shown
How do you get started with Mirai?
- 01Open trymirai.com and read the models and runtime sections
- 02Review the uzu engine and the Qwen, LFM and Muse-Glimmer model details
- 03Use the contact form to ask about access, the Mac app and the SDK
Mirai: frequently asked questions
What does Mirai Labs actually build?
Mirai Labs builds uzu, a Rust LLM engine for iPhone, iPad and Mac, along with 4-bit and 8-bit Qwen, LFM and Muse-Glimmer models, plus an application layer for Apple device automation.
Does Mirai run in the cloud or on the device?
The stated focus is on-device inference on Apple Silicon. The site describes on-device as a different system from cloud serving, with batch size 1, latency-first execution and constrained memory.
How much does Mirai cost?
Pricing is not published. The site routes interested visitors to a contact page with the message "Talk to us".
How do the models compare with other runtimes?
The meta description states that the Qwen, LFM and Muse-Glimmer models are benchmarked against MLX and llama.cpp, though the source does not include the benchmark numbers.
Mirai alternatives
The closest tools in our directory, matched on category, focus areas and audience.
| Tool | Best for | Free plan | Starting price | Focus areas |
|---|---|---|---|---|
| Mirai | Developers wanting low-latency local LLM inference on Apple devices | No | — | On Device |
| TensorFlow Hub | TensorFlow developers reusing and fine-tuning pre-trained models | Unknown | — | Fine Tuning, On Device |
| CloudflareAI | Low-latency execution of ML models worldwide. | Unknown | — | — |
| customvision.ai | Customize computer vision models with your own images and REST APIs. | Unknown | — | — |
| Fal.ai | Generative media platform for developers to run diffusion models with fast AI inference. | Unknown | — | — |
| Flyte V1 3 | Flyte is a scalable workflow automation platform for data and ML processes. | Unknown | — | — |
| Nyckel | Image recognition for retail inventory management. | Unknown | — | — |
| Ratio1 | Decentralized AI OS powered by blockchain for rapid AI app development and deployment. | Unknown | — | — |
| TunemyAI | Optimize and deploy AI models effortlessly with Tune My AI, the leading tool for AI model tuning and deployment. | Unknown | — | — |
| NBULA | MENA enterprises and agencies needing Arabic-first, in-region AI deployment | No | — | — |
| Applied Compute | Enterprises building their own models | No | — | — |

