Mirai Labs
Mirai Labs is an on-device AI lab building models, an inference runtime, and a quantization/optimization stack for real-time, private local intelligence on Apple Silicon. Its vision is to make AI an ambient interface that can execute tasks locally rather than relying on isolated prompts, app switching, or cloud inference.
At a Glance
- AI model companies and frontier model providers
- AI research labs and researchers
- Independent developers and makers
- Consumer application developers
- +2 more
AI Tools by Mirai Labs
(1)
uzu
On-Device LLM Engine for Apple
Discussions
No discussions yet
Be the first to start a discussion about Mirai Labs
Latest News
The Computer Should Know the Way: Mirai outlined its vision for AI-native, on-device computer interaction.
Tackling random foreign characters: Mirai explained and fixed a sampling issue in uzu/speculative decoding outputs.
Speculative decoding in uzu: Mirai published details of its draft-and-verify decoding implementation.
Sparse Buffers for KV Cache: Mirai described its Metal sparse-buffer approach for memory-efficient on-device inference.
Products & Services
Open-source Rust inference engine using Metal and Apple unified memory to run AI models locally in apps and expose high-level APIs across Rust, Python, Swift, and TypeScript.
Open-source model conversion and optimization toolkit for bringing popular and custom model architectures to Apple devices; supports quantization, validation, and speculative-decoding draft-model workflows.
A library of Apple-device-optimized checkpoints, including Qwen and LiquidAI families, with Mirai-M 4-bit and Mirai-L 8-bit quantization variants and per-device benchmarks.
A native macOS and Apple Silicon chat application positioned as a private, faster alternative to Ollama and LM Studio, with local model downloads and support for Qwen, Liquid AI, and Muse-Glimmer models.
Market Position
Mirai positions itself as a full-stack, Apple-Silicon-specific alternative to generic inference runtimes: it co-designs models, quantization, and a Rust/Metal runtime rather than wrapping MLX, llama.cpp, or other general-purpose stacks. Its direct competitive set includes MLX, llama.cpp, Ollama, LM Studio, and other edge-inference providers; it differentiates on single-request latency, memory efficiency, quantization quality, and local privacy. It has also described future Android support, while its current product focus is Apple Silicon.
Leadership
Founders
Dima Shvets
Co-founder and co-CEO; launched/co-founded Reface, a generative photo/video AI company that reached more than 300 million users and was backed by Andreessen Horowitz; later became an a16z scout.
Alexey Moiseenkov
Co-founder and co-CEO; co-founded and led Prisma, an on-device AI photo-enhancement company that reached more than 100 million monthly active users and won App of the Year recognition on the App Store and Google Play.
Executive Team
Dima Shvets
Co-founder and co-CEO
Former Reface co-founder/executive; built generative photo/video AI products used by more than 300 million people.
Alexey Moiseenkov
Co-founder and co-CEO
Former Prisma co-founder/CEO; built an early on-device neural-network photo-enhancement product used by more than 100 million monthly active users.
Board of Directors
Founding Story
Mirai was founded in London in 2024 by the entrepreneurs behind Reface and Prisma after they saw that the generative-AI conversation focused heavily on cloud servers while consumer on-device AI remained the missing piece. They set out to build a complete, hardware-aware stack—models, runtime, and quantization—from the device constraint up, initially targeting Apple Silicon.
Business Model
Revenue Model
The open-source uzu and lalamo libraries are free under the MIT license. Mirai also offers commercial model conversion, optimization, benchmarking, and deployment support for model companies, researchers, and developers, with sales initiated through a contact/demo process; its hybrid device/cloud stack can support cloud inference alongside local execution.
Pricing Tiers
Both libraries are documented as fully open source under the MIT license.
Target Markets
- AI model companies and frontier model providers
- AI research labs and researchers
- Independent developers and makers
- Consumer application developers
- Companies seeking lower-cost, lower-latency, privacy-preserving edge inference
- Apple ecosystem users and developers
- Embedding private local AI assistants and chat into consumer apps
- Running text and voice models on iPhone, iPad, and Mac without network latency
- Model companies distributing and benchmarking models on Apple hardware
- AI researchers and labs testing real-world edge performance
- Independent developers fine-tuning or training models and deploying them locally
- Hybrid AI applications that route unsuitable workloads from device to cloud