Mirai

What Mirai does

Mirai Labs (operating as Mirai Tech Inc.) is a London-headquartered on-device AI infrastructure company building a full-stack inference runtime for Apple Silicon. Its core asset is uzu, a from-scratch Rust inference engine co-designed with a proprietary quantization scheme (Mirai-M 4-bit and Mirai-L 8-bit) and a Metal placement sparse-buffer KV cache subsystem, optimized for batch-size-1, latency-first, memory-constrained execution on iPhone, iPad and Mac. The company claims 37% faster generation and 59% faster pre-fill versus Apple's MLX and 40–60% more tokens per second than Unsloth/llama.cpp and MLX at comparable quality, and ships a curated local model library that bundles Mirai-quantized Qwen3.5 checkpoints with partner models from Liquid AI, Google, Meta and OpenAI.

Mirai monetizes through two channels: a freemium self-serve layer consisting of the free Mirai for Mac desktop app (positioned as a faster alternative to Ollama and LM Studio) and the open-source uzu SDK distributed across Cargo, Swift Package Manager, NPM and PyPI; and a quote-based enterprise motion targeting model companies, AI labs and chip manufacturers that want their weights optimized, benchmarked and distributed on Apple devices, with conversion/quantization handled by the lalamo toolkit. Its primary customer segment is foundation-model providers seeking to reach the Apple installed base without building in-house on-device infrastructure.

The company was founded by Dima Shvets (Reface, 300M users) and Alexey Moiseenkov (Prisma, 100M MAU), is advised by Apple MLX co-creator Awni Hannun, and closed a $10M seed round in February 2026 led by Uncork Capital with participation from former Stripe CTO David Singleton and Snowflake co-founder Marcin Żukowski. As of mid-2026 the team stands at roughly 16 people distributed across London, San Francisco and remote Europe, and the stated roadmap is to extend the stack beyond Apple Silicon into multimodal (voice, vision) workloads and Android via chip-manufacturer partnerships.

Name

Mirai

Legal name

Mirai Tech Inc.

Website

https://trymirai.com

Company type

Private

Founded year

2024

Operating status

Operating

Headcount range

1–10 employees

Short description

Mirai Labs builds uzu, a from-scratch Rust on-device AI inference engine and co-designed quantization stack for Apple Silicon, sold to model companies and AI labs and delivered to developers via an open-source SDK and a free macOS app.

Ownership category

akta.pro rank

Product category

On-Device AI Inference Software

NAICS

Computer Systems Design Services (541512), Computer Systems Design and Related Services (54151), Computer Systems Design and Related Services (5415)

SIC

Services-Computer Programming Services (7371), Services-Computer Integrated Systems Design (7373)

akta.pro primary industry

Model Serving, Inference & Deployment Platforms (APIs, Edge/On-Prem) (HDAEANAC)

akta.pro secondary industries

LLMOps & Generative AI Platforms (Prompt/Agent Orchestration, RAG) (HDAEANAD), Model Hosting, Serving & Inference Platforms (HDAAACAB), Enterprise Foundation Model Integration & APIs (Connectors, Governance, Deployment) (HDAAACAO), On-Device Inference Runtimes & SDKs (mobile/embedded) (HDAAAJAB), Model Deployment, Serving & Inference Platforms (HDAAABAF)

Keywords

• On-device AI inference
• LLM inference engine
• Apple Silicon optimization
• Model quantization software
• AI developer SDK

Headquarters

HQ city

Hermosa Beach

HQ country

United States

HQ region

North America

Offices3 records

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Marketing or Sales, Operations

Revenue model

1. macOS desktop app distribution: Native Apple Silicon macOS app distributed via direct download from trymirai.com and via the 'mirai' Homebrew package; positioned as a faster, simpler alternative to Ollama and LM Studio. Pricing is not publicly disclosed.
2. Enterprise inference engine / SDK licensing: Quote-based B2B engagement for the uzu inference engine and conversion tooling, targeting model companies, AI labs, frontier-model providers and chip manufacturers that want their models optimized, benchmarked and distributed on Apple devices. Pricing is via the 'Talk to us' contact form rather than public tiers.
3. Open-source SDK community release: The uzu Rust inference engine and language bindings (Swift, Python, Node.js, Rust) are released as open-source on GitHub, building adoption and ecosystem presence as a funnel into paid enterprise engagements.

Pricing tiers

Go-to-market motion4 records

Distribution channels4 records

Marketing channels6 records

Core offering

Mirai builds and licenses an end-to-end on-device AI stack for Apple Silicon, anchored by uzu — a Rust-based, hardware-aware LLM inference engine — combined with a co-designed quantization scheme (Mirai-M 4-bit and Mirai-L 8-bit), a multi-language developer SDK, a one-command model conversion toolkit, and a curated local model library. The stack is distributed via a free native macOS app for individual developers and via quote-based enterprise engagements with AI labs, foundation-model providers, and chip manufacturers that want their models optimized, benchmarked, and shipped to the Apple device installed base.

Product overview

Mirai Labs (Mirai Tech Inc.) is a frontier on-device AI lab that offers an integrated stack — models, inference runtime, and quantization — rather than a single product. The core of the stack is uzu, a Rust-based inference engine optimized for Apple Silicon that delivers up to ~38% faster prompt processing than Apple's MLX and supports speculative decoding, sparse-buffer KV cache paging, and structured output. Developers embed uzu into their own apps via the Mirai SDK, which is distributed in Rust (Cargo), Swift (Swift Package Manager), TypeScript/Node.js (NPM), and Python (PyPI), with Kotlin planned; a Platform portal at platform.trymirai.com and docs at docs.trymirai.com accompany it. End users can run models through the Mirai for Mac desktop app (a native Apple-Silicon alternative to Ollama and LM Studio) and browse ready-to-run weights in the Mirai Models Library, which bundles Mirai's own quantized Qwen3.5 checkpoints (Mirai-M / Mirai-L), Liquid AI LFM2 / LFM2.5 partner models, and community MLX-format conversions. Model providers and developers use the Mirai Conversion and Optimization Toolkit (lalamo) to convert and optimize their own checkpoints for on-device deployment with one command.

Differentiator

Problem solved

Functional benefit

Brands

• uzu: Mirai's open-source high-performance LLM inference engine built in Rust for on-device inference on Apple Silicon; distributed via Cargo, Swift Package Manager, NPM, and PyPI.

Products and services

• uzu Inference Runtime (Inference Engine) Rust-based, hardware-aware on-device LLM inference engine targeting Apple Silicon (iPhone, iPad, Mac), described as the fastest inference runtime for those devices. Powers structured output, task-specific sessions, speculative decoding, sparse-buffer KV cache paging, and built-in performance metrics for foundation-model providers and AI labs that want to ship models natively on Apple hardware.
• Mirai SDK (uzu SDK) Cross-language SDK that wraps the uzu Rust inference core so developers can embed on-device AI inference into their own Apple-platform applications. Distributed via Cargo (Rust), Swift Package Manager (Swift), NPM (TypeScript/Node.js), and PyPI (Python), with Kotlin support marked Coming Soon, providing a unified high-level API across all supported languages.
• Mirai for Mac (Chat for Mac) Native macOS / Apple Silicon desktop application for running local LLMs with complete on-device privacy, positioned as a faster and simpler alternative to Ollama and LM Studio. Supports Gemma, Qwen, Llama, DeepSeek, Hugging Face-sourced models, and 'Polaris', targeting individual developers and Apple-platform end users.
• Mirai Models Library (Local Models) Curated, on-device-optimized library of AI models available for instant deployment through Mirai's runtime. Includes Mirai's own quantized checkpoints (Qwen3.5-0.8B/2B/4B in Mirai-M and Mirai-L), partner models from Liquid AI (LFM2, LFM2.5), and community MLX-format conversions (mlx-community), spanning 350M–27B parameters across Qwen, Gemma 3, Llama 3.2, GPT-OSS-20B, and LFM2/LFM2.5.
• Mirai Conversion and Optimization Toolkit (lalamo) Command-line toolkit (lalamo) that converts and optimizes third-party or fine-tuned LLMs for on-device inference on iPhone, iPad, and Mac in one command. Automatically fetches models from Hugging Face, supports popular architectures out of the box (Qwen, Gemma, Llama, LFM, GPT-OSS), handles custom architectures via config or hand-built converters, trains draft models for speculative decoding, applies Mirai's quantization scheme, and validates output correctness layer-by-layer.

Quantifiable outcome

• 37% faster generation and 59% faster pre-fill compared with Apple's MLX on Apple Silicon (uzu inference engine)
• +5 more outcomes

Segments5 records

Ideal customer profiles4 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration1 record

AI capability12 records

Feature8 records

Scale indicators10 records

Recent moves6 records

Expansion highlights6 records

Broad incumbents

• CoreWeave: CoreWeave is a large-scale GPU cloud infrastructure provider serving AI inference and training workloads. It represents the cloud-side alternative to on-device inference and competes for the same enterprise AI workloads via a different cost/latency profile.
• Groq: Groq provides high-speed inference infrastructure (cloud-based LPU hardware). While not on-device, it competes for the same customer demand — replacing slow or expensive cloud inference — that Mirai addresses via local execution on Apple Silicon.
• Hugging Face: Hugging Face is the dominant model hub and ecosystem player whose repositories Mirai's runtime pulls from and whose distribution Mirai uses for its own quantized checkpoints. They overlap in serving the open-weight model ecosystem, though HF is a much broader platform.

Emerging players

• Liquid AI: Liquid AI produces the LFM2 / LFM2.5 small on-device models featured in Mirai's model library. They are simultaneously a partner (models distributed via Mirai) and a potential competitor with their own on-device inference ambitions.

Direct peers

• llama.cpp: llama.cpp is the foundational open-source C/C++ LLM inference engine that most local runners (Ollama, LM Studio) wrap. Mirai benchmarks its uzu engine directly against llama.cpp and competes for the same developer mindshare in on-device inference.
• Unsloth: Unsloth provides optimized fine-tuning and inference for LLMs with strong open-source adoption. Mirai's quantization benchmarks compare directly against Unsloth (and llama.cpp), positioning both as competing open-source performance optimizers.
• LM Studio: LM Studio is a popular cross-platform desktop app for running local LLMs, including on Apple Silicon. Mirai's Chat for Mac app is explicitly framed as a faster native-Apple-Silicon alternative to LM Studio.
• Ollama: Ollama is the most widely used local LLM runner for macOS and other platforms. Mirai explicitly positions its macOS app as a faster, simpler alternative to Ollama, making it the closest direct competitor in the on-device inference consumer app category.
• Apple MLX: Apple's first-party ML framework optimized for Apple Silicon. Mirai's headline value proposition — 37–59% speedups over MLX — is defined in direct comparison with this framework, making it both a competitor and the platform incumbent Mirai is differentiated against.

Others

• Anthropic: Anthropic is a frontier foundation-model provider whose on-device strategy (and whose employee Awni Hannun, an Apple MLX co-creator, advises Mirai) is adjacent to Mirai's distribution business. They are an ecosystem partner rather than a direct competitor.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights6 records

Customer concentration

• Xx.com/getmirai
• GitHubgithub.com/trymirai
• LinkedInlinkedin.com/company/trymirai

Revenue estimate

Valuation estimate

Number of profiles

Profiles3 records

Funding overview

Funding rounds2 records

Investors6 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Mirai