M

Mobilellm 125M

Developed by facebook
MobileLLM is Meta's series of sub-billion parameter language models specifically optimized for resource-constrained devices, significantly improving on-device inference efficiency through deep-narrow architecture design.
Downloads 1,675
Release Time : 10/30/2024

Model Overview

This model series adopts an optimized Transformer architecture designed for mobile devices and edge computing scenarios, outperforming models of similar parameter scales in tasks like common sense reasoning.

Model Features

Device-optimized architecture
Employs deep-narrow structure design with embedding sharing and grouped query attention (GQA), significantly reducing memory footprint
Efficient inference performance
The 125M version achieves 3.7% higher accuracy in common sense reasoning tasks compared to the previous OPT-125M
Parameter scalability
Offers multiple parameter scales from 125M to 1.5B to accommodate different hardware conditions

Model Capabilities

Text generation
Common sense reasoning
On-device deployment

Use Cases

Mobile applications
Smart keyboard prediction
Enables low-latency text input prediction on mobile devices
The 125M model achieves real-time inference on mid-range mobile chips
Educational tools
Offline learning assistant
Provides Q&A functionality for educational devices in offline environments
The 350M version achieves 52.1% accuracy on the ARC-Challenge test set
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
Ā© 2025AIbase