← Retour au blog
tech 5 August 2026

Maple-Preview: AI on iPhone at Unprecedented Speed

Discover how Maple-Preview leverages a Ternary 20B MoE model to deliver a performance of 120 tok/s on a simple iPhone. A revolution for mobile AI.

Article inspired by the original source
Show HN: Maple-Preview – Ternary 20B MoE running at 120 tok/s on a iPhone ↗ deepgrove.ai

Introduction

In the world of artificial intelligence, every technological advance can disrupt the way we design and interact with machines. Today, we delve into a major innovation: Maple-Preview, which allows a Ternary 20B MoE model to run at an impressive speed of 120 tok/s directly on an iPhone. Why is this a significant breakthrough? Let's explore.

What is Maple-Preview?

Maple-Preview is a tech showcase by DeepGrove, a pioneer in advanced artificial intelligence. This project highlights a Ternary 20B Mixture of Experts (MoE) model, an architecture that uses multiple experts to optimize data processing. The innovation here lies in running this model on a device as common and accessible as an iPhone.

The Technology Behind Maple-Preview

Ternary Models

Ternary models use three possible states to store and process information, unlike traditional binary models that use only two. This allows for a significant reduction in model size while maintaining (or even improving) efficiency.

Mixture of Experts (MoE)

The MoE's peculiarity is its ability to activate only a portion of the model during each prediction. This means millions of parameters can be used without requiring exorbitant computational power. This model is particularly well-suited to be run on mobile devices like the iPhone.

Impressive Performance

120 tok/s

One of Maple-Preview's most impressive aspects is its ability to process 120 tok/s. To put this in perspective, many state-of-the-art AI models on powerful servers barely reach this level of performance. Achieving this on an iPhone is a technical feat.

Cost Reduction

The ability to deploy complex models on everyday devices significantly reduces infrastructure costs. Companies can thus offer high-performance AI solutions without investing in expensive servers.

Potential Use Cases

Mobile Applications

Imagine real-time translation apps, voice command applications, or data analysis directly on your mobile. Maple-Preview paves the way for advanced features accessible to everyone.

Edge Computing

With this technology, edge computing becomes a tangible reality. Data can be processed locally, thus reducing latency and improving privacy.

Conclusion

Maple-Preview is not just a technical demonstration; it's a revolution in how we envision artificial intelligence on mobile devices. By democratizing access to high-performance models, DeepGrove is transforming AI into a truly accessible tool. Let's discuss your project in 15 minutes.

Maple-Preview Ternary MoE AI on iPhone DeepGrove edge computing
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call