Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Octopus v3: Technical Report for On-device Sub-billion Multimodal AI Agent

Wei Chen, Zhiyuan Li

TL;DR

Octopus v3 presents a sub-billion-parameter multimodal agent optimized for edge devices, capable of processing English and Chinese and operating on hardware as constrained as a Raspberry Pi. It fuses CLIP-based visual encoding with a novel functional-token mechanism and a multi-stage training pipeline to convert images and language into actionable on-device tokens. In evaluations against GPT-4V/GPT-4, the model with under 1B parameters achieves competitive performance on ten smartphone-API tasks, highlighting strong edge-deployment potential and scalable token-based customization. The work discusses social impact, licensing, and avenues for expanding modalities, outlining a path toward private, on-device AI agents with broad real-world applicability.

Abstract

A multimodal AI agent is characterized by its ability to process and learn from various types of data, including natural language, visual, and audio inputs, to inform its actions. Despite advancements in large language models that incorporate visual data, such as GPT-4V, effectively translating image-based data into actionable outcomes for AI agents continues to be challenging. In this paper, we introduce a multimodal model that incorporates the concept of functional token specifically designed for AI agent applications. To ensure compatibility with edge devices, our model is optimized to a compact size of less than 1B parameters. Like GPT-4, our model can process both English and Chinese. We demonstrate that this model is capable of operating efficiently on a wide range of edge devices, including as constrained as a Raspberry Pi.

Octopus v3: Technical Report for On-device Sub-billion Multimodal AI Agent

TL;DR

Octopus v3 presents a sub-billion-parameter multimodal agent optimized for edge devices, capable of processing English and Chinese and operating on hardware as constrained as a Raspberry Pi. It fuses CLIP-based visual encoding with a novel functional-token mechanism and a multi-stage training pipeline to convert images and language into actionable on-device tokens. In evaluations against GPT-4V/GPT-4, the model with under 1B parameters achieves competitive performance on ten smartphone-API tasks, highlighting strong edge-deployment potential and scalable token-based customization. The work discusses social impact, licensing, and avenues for expanding modalities, outlining a path toward private, on-device AI agents with broad real-world applicability.

Abstract

A multimodal AI agent is characterized by its ability to process and learn from various types of data, including natural language, visual, and audio inputs, to inform its actions. Despite advancements in large language models that incorporate visual data, such as GPT-4V, effectively translating image-based data into actionable outcomes for AI agents continues to be challenging. In this paper, we introduce a multimodal model that incorporates the concept of functional token specifically designed for AI agent applications. To ensure compatibility with edge devices, our model is optimized to a compact size of less than 1B parameters. Like GPT-4, our model can process both English and Chinese. We demonstrate that this model is capable of operating efficiently on a wide range of edge devices, including as constrained as a Raspberry Pi.
Paper Structure (19 sections, 1 figure, 10 tables)

This paper contains 19 sections, 1 figure, 10 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Methodology
  4. Encoding visual information
  5. Functional token
  6. Multi-stage training
  7. Model evaluation
  8. Send email
  9. Send text message
  10. Google search example
  11. Amazon purchase example
  12. Smart recycle
  13. Lost and found
  14. Interior design
  15. Instacart shopping
  16. ...and 4 more sections

Figures (1)

  • Figure 1: Sub-billion multimodal AI agent with vision