Mirai Secures $10M Seed Funding to Advance On‑Device AI Model Inference

Why This Funding Matters for Mobile AI

In a landscape dominated by cloud‑centric AI services, Mirai’s focus on running sophisticated models directly on phones, laptops, and other consumer hardware addresses a critical bottleneck: latency, privacy, and cost. The fresh capital will enable the team to expand its engineering roster, broaden platform support beyond Apple Silicon, and bring a developer‑friendly SDK to market by the end of 2026.

For readers interested in broader AI startup trends, our AI news hub offers daily updates on breakthroughs and funding rounds.

Founders’ Track Record: From Reface to Prisma

The duo behind Mirai—Dima Shevts and Alexey Moiseenkov—have already proven their ability to build viral consumer AI products at scale.

• Dima Shevts: Co‑founder of Reface, the face‑swapping app that attracted a16z backing and amassed over 100 million downloads. After exiting Reface, Shevts served as a scout for a16z, deepening his network in venture capital.
• Alexey Moiseenkov: Former CEO and co‑founder of Prisma, the AI‑powered photo‑filter platform that went viral worldwide and was later acquired by a major tech conglomerate.

Both founders have spent the last decade engineering consumer‑grade AI pipelines that run efficiently on limited hardware. Their experience gave them a front‑row seat to the growing frustration among developers: “Why are we still paying per‑token cloud fees for tasks that could run locally on a phone?” they asked.

Mirai’s Core Technology: Edge‑Optimized Inference Engine

Mirai’s flagship product is a Rust‑based inference engine that sits between a model and the device’s hardware accelerator. The engine delivers three concrete benefits:

1. Speed Boost: Benchmarks show up to a 37 % reduction in latency for text generation tasks on Apple Silicon.
2. Zero‑Loss Quantization: The engine optimizes execution without altering model weights, preserving output quality.
3. Developer Simplicity: An SDK that requires fewer than ten lines of code to enable summarization, classification, or translation directly on the device.

The current release targets text and voice modalities, with a roadmap that includes vision models later in 2027. Mirai is also building an orchestration layer that intelligently routes heavyweight requests to the cloud while keeping latency‑sensitive operations on‑device.

The company’s approach aligns with the emerging “edge‑first” paradigm championed by chip makers such as Qualcomm and Apple, who are investing heavily in on‑device neural engines.

The $10 Million Seed Round: Who’s Betting on Edge AI?

The round was led by Uncork Capital, a firm with a history of backing early‑stage machine‑learning infrastructure startups. Notable participants included:

The diverse investor mix—spanning venture capital, enterprise SaaS, and AI‑focused founders—signals strong confidence that on‑device inference will become a core cost‑saving lever for consumer apps.

For a deeper dive into startup financing trends, explore our startup funding insights.

Market Impact: How Mirai Could Redefine Mobile AI

The economics of cloud inference are shifting. According to a recent Forrester report, enterprises spend an average of $0.12 per 1,000 tokens on cloud AI services—a cost that scales dramatically with user growth. By moving inference to the edge, Mirai promises to cut these expenses by up to 70 % for high‑frequency workloads.

Potential use‑cases include:

• On‑device personal assistants that answer queries without sending data to the cloud.
• Real‑time transcription and translation for travelers, preserving privacy.
• AI‑enhanced camera apps that apply complex filters instantly.
• Gaming AI that reacts locally, eliminating lag.

Mirai’s roadmap for 2026‑2027 includes:

• Q3 2026: Public beta of the SDK for iOS developers.
• Q4 2026: Android support and integration with Google’s Tensor Processing Units.
• Q2 2027: Vision model acceleration (image classification, object detection).
• Q4 2027: Marketplace of pre‑tuned edge models, enabling developers to plug‑and‑play.

The upcoming marketplace aligns with UBOS’s own template ecosystem, where developers can quickly spin up AI‑powered applications using pre‑built components.

Architecture Snapshot

Mirai’s stack consists of three layers:

1. Model Adapter: Converts ONNX or TensorFlow Lite models into a unified intermediate representation.
2. Runtime Engine (Rust): Executes the model using SIMD, Metal (iOS), or Vulkan (Android) primitives, applying operator‑level graph optimizations.
3. Orchestration Layer: Monitors device resources and decides whether to run locally or forward to a cloud endpoint.

The engine’s “zero‑loss” claim stems from a dynamic precision scheduler that selects the highest‑precision path that satisfies a target latency budget, avoiding the quality degradation typical of static quantization.

How Mirai Stands Apart from Competitors

While Apple’s Core ML and Google’s TensorFlow Lite provide on‑device inference, they require developers to manually tune models and often sacrifice performance. Mirai differentiates itself by:

• Automatic Optimization: One‑click SDK integration with built‑in performance profiling.
• Cross‑Platform Consistency: Same runtime works on iOS, macOS, and Android, reducing engineering overhead.
• Edge‑First Marketplace: Future marketplace of pre‑optimized models, similar to UBOS’s partner program for AI services.