Skip to content

MobileLLM-600M

A 600M-parameter language model built for on-device use

Visit official site

What MobileLLM-600M is

MobileLLM-600M is described as an autoregressive language model from Meta with an optimised Transformer architecture for devices with limited resources. Its design extends to larger variants such as MobileLLM-1B and MobileLLM-1.5B.

What you can do with it

  • Run zero-shot commonsense reasoning tasks
  • Load pretrained weights from the HuggingFace platform
  • Fine-tune or evaluate the model with the provided code examples
  • Pick a size from 125M to 1.5B parameters

Application scenarios

  • Run text generation and understanding on a mobile device
  • Serve as the back-end model for a chatbot
  • Add voice interaction to smart home devices

Getting started

  • Search for the model on the HuggingFace site
  • Load pretrained weights with the code examples provided
  • Add special tokens per the HuggingFace guide for tuning
  • Clone the MobileLLM GitHub codebase and install its dependencies

Pros & cons

✓ What we like

  • Pretrained weights load from the HuggingFace platform
  • The MobileLLM codebase ships pretraining code

! What to watch out for

  • Built around on-device, resource-constrained settings
  • Fine-tuning and evaluation follow HuggingFace guide steps

FAQ

Who developed it?

The material names Meta as the developer of the autoregressive model.

What accuracy gains are reported?

The material states 2.7% over 125M and 4.3% over 350M models on zero-shot commonsense tasks.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review