</aside>

<aside> <img src="notion://custom_emoji/1033d2c0-f40d-4c2f-9583-313aa8d2b337/1a080ee7-e5ee-8009-889f-007a7a44cbfb" alt="notion://custom_emoji/1033d2c0-f40d-4c2f-9583-313aa8d2b337/1a080ee7-e5ee-8009-889f-007a7a44cbfb" width="40px" /> MIZU is the first Edge AI Data Network, turning personal devices into a self-hosted AI agent ecosystem. Powered by local models, MIZU’s AI agents autonomously manage, process, and share personalized data across apps like Telegram and Twitter—securely and seamlessly

</aside>

MIZU Overview

Unlocking the Potential of Edge AI

Edge AI is 5–10x larger than traditional data centers, with a compute capacity of 20–40 ZFLOPS compared to just 4 ZFLOPS in centralized infrastructure. With the rise of self-hosted AI models via platforms like Ollama and LMStudio, individuals are increasingly running AI locally.

However, turning edge devices into a functional AI inference network remains a challenge due to unreliable nodes, execution verification issues, and fragmented compute power. Addressing these barriers is critical to unlocking the full potential of decentralized AI.

The Challenges of Edge AI Inference

1. Unreliable Nodes: The Availability Problem

Unlike cloud servers, which operate under strict uptime guarantees, edge devices are inherently unreliable. Home computers and personal GPUs can be turned off, go to sleep, or disconnect at any moment, making it difficult to ensure a consistent inference service.

Various approaches have been proposed to mitigate this, most commonly through request redundancy, where multiple devices handle the same query, and only the fastest valid response is accepted. While this method can improve reliability, it introduces significant inefficiencies by duplicating compute work.

Another approach focuses on smart request routing, where tasks are dynamically assigned to devices with a history of stable uptime. Additionally, economic incentives are being explored, rewarding nodes that maintain high availability while penalizing those that frequently drop offline.

2. Execution Verification: Trusting the Computation

One of the biggest risks in a decentralized AI network is the ability for bad actors to cheat the system. If a worker can claim to have performed inference but return incorrect or empty results, the network loses integrity.

Current approaches to execution verification fall into three categories:

Optimistic challenge models – Users can challenge inference results if they suspect fraud, penalizing dishonest workers. This method, however, relies on active participation and is not foolproof.
Per-worker sampling – Randomized checks verify a subset of worker outputs, balancing overhead and trust. The challenge lies in determining the right sample rate to minimize cost while maintaining accuracy.
KV cache tracking – An experimental approach that analyzes memory states to confirm inference execution. This is still an evolving field with practical deployment challenges.

3. Fragmented Compute Power: Managing the Hardware Mess

Unlike cloud environments with standardized hardware, Edge AI networks must deal with a wide range of devices—from high-performance desktops to low-power mobile phones. Managing this diversity is a logistical challenge, especially when some devices can barely run small models, while others can handle large-scale inference.

One approach is to move edge devices into controlled environments, such as micro data centers, and limit the number of supported hardware SKUs (e.g., Mac Mini or Intel NUCs). While this simplifies hardware management, it fundamentally contradicts the purpose of Edge AI, centralizing compute power rather than distributing it.

Alternatively, some research explores federated inference, where workloads are distributed across multiple weaker devices to enable larger model execution on edge hardware. While this approach makes it possible to run bigger models, individual device latency remains high, and throughput is significantly limited, reducing overall efficiency.

Lessons from MIZU: Phones vs. Laptops

In Phase 1, MIZU focused on integrating smartphones into the network. While some phones can run 1–3B parameter models, we shifted our focus to laptops and desktops in Phase 2 due to two critical challenges:

iPhones aggressively terminate background processes, making it nearly impossible to maintain a persistent AI node—even within a dedicated app.
Android devices struggle with performance and GPU support. While keeping an Android device awake is possible, most lack the processing power to run inference efficiently. Even high-end Android devices with capable hardware suffer from poor GPU optimization, leading to inconsistent model execution and inefficient power usage.

Although we continue to explore ways to integrate smartphones, our immediate focus is on devices that ensure network stability and reliability. To achieve this, MIZU prioritizes devices that meet two key criteria:

Sufficient performance to run AI models effectively.
Sustained uptime, ensuring that inference nodes remain available for long periods

MIZU’s Approach to Building the Edge AI Network

MIZU is building an edge AI inference network optimized for asynchronous workloads, ensuring scalability, efficiency, and cost-effectiveness.

1. Asynchronous Inference First

Unlike cloud-based models that prioritize low-latency responses, MIZU is designed for large-scale batch inference and data processing, where cost and scalability matter more than real-time latency.

MIZU functions as both a routing and caching layer, reducing redundant processing while ensuring reliability across distributed edge devices.

2. Optimistic Challenge + Per Worker Sampling: The Progressive Reputation System

MIZU employs a hybrid verification approach that combines optimistic challenges and per-worker sampling, ensuring worker execution is trustworthy while maintaining efficiency. Each MIZU Pool owner can enable per-worker sampling to enhance result reliability. Over time, workers with consistent performance will require fewer checks, reducing network overhead.

Reputation-Based Sampling Strategy

The default sampling strategy defines a reputation score (r) for each worker:

c = count of correctly finished inferences
c’ = count of incorrect inference results
c’’ = count of inference results not returned in time
r = c - k1 * c’ - k2 * c’’, where k1 = 10, k2 = 5 (default penalties)

The sampling rate (s) is calculated as s = Max(1, r) / 10,000

With this approach, a well-behaving worker will see its sample rate drop to 0.01% after 200 correct inferences, ensuring minimal verification overhead while maintaining execution integrity.

MIZU is continuously refining verification strategies, with plans to introduce more dynamic reputation models in future updates.

3. Pool Ownership: Competition Drives Quality

MIZU allows anyone to own and operate a MIZU Pool, where pool owners:

Earn a share of revenue based on the activity of their network.
Onboard and manage new devices to expand compute power.
Ensure quality control using MIZU’s built-in monitoring infrastructure.

This market-driven model encourages efficiency and network stability while distributing decision-making across multiple stakeholders.

4. Application-Driven Growth: The Open Content Aggregator

Most decentralized AI networks struggle due to a lack of real-world demand. MIZU addresses this by launching its first major application: The Open Content Aggregator—a tool that collects, processes, and redistributes content from Telegram, Twitter, and other decentralized data sources.

This aggregator runs on periodic AI pipelines, making it a perfect fit for MIZU’s async inference model, where:

Traders can track KOL insights across Telegram and Twitter.
Ecosystems can compile developer Q&As into evolving FAQs.
KOLs and analysts can aggregate industry reports into structured content.

By focusing on real demand, MIZU ensures steady AI workload distribution, fueling network adoption.

Companion Computing: The Bridge to Adoption

Instead of relying solely on edge devices from day one, MIZU follows a hybrid model to ensure reliability while transitioning to full decentralization:

AI workloads start on traditional GPU networks.
1. Over time, tasks shift to the Edge AI network to reduce costs.
1. If needed, workloads fall back to centralized GPUs for reliability.

This gradual transition ensures businesses can lower costs while maintaining high performance.

Looking Ahead: Edge AI Hardware & Future Expansion

While software solutions drive current progress, specialized AI hardware could accelerate edge inference adoption. MIZU is exploring custom AI workstations, designed to function as decentralized inference nodes, offering

Optimized hardware for AI workloads—balancing CPU, GPU, and storage.
Always-online functionality, ensuring stable network participation.

These explorations could redefine how AI inference is performed at the edge, further reducing reliance on centralized infrastructure.