Skip to main content

​
Overview

​
Prerequisites

Before you begin, please ensure you have the following set up:
  • The aos command-line interface (CLI): This is the primary tool for interacting with the AO ecosystem. If you don’t have it installed, open your terminal and run: 
    npm i -g https://preview_ao.arweave.net
    If you’re new to aos or want a refresher, we highly recommend reading the official AO Cookbook guide on connecting aos with HyperBEAM nodes.
  • An Arweave Wallet: You’ll need a wallet (like ArConnect or Wander) to create your aos process. Your process ID is your identity on the network and is required to receive credits.

​
Key Concepts & Terminology

To use our service effectively, you’ll need to understand these two core concepts:
  • HyperBEAM vs. Legacy Network The AO ecosystem consists of two main environments: the legacynet and the newer, high-performance HyperBEAM. Our APUS AI Service is built exclusively on HyperBEAM to leverage its unique capabilities for GPU integration.
  • Router Process The Router Process is the “front door” to the APUS AI service. It’s a single, stable AO process that you will send all your inference requests to. It manages the request queue, validates payments, and dispatches tasks to available GPU workers. You will use this Process ID as the Target for all your interactions.
    • Router Process ID: TED2PpCVx0KbkQtzEYBo0TRAO-HPJlpCMmUzch9ZL2g
      The previous router Bf6JJR2tl2Wr38O2-H6VctqtduxHgKF-NzRB9HhTRzo is still active but experiencing backpressure.
  • Credits Credits are the units of payment for using the APUS AI service. Every inference request consumes a fixed amount of credits.

​
Service Capabilities & Hackathon Limitations

To ensure a smooth and predictable experience for all participants during the Hackathon, our service is operating with the following specifications.
  • AI Model: Gemma-3-12B (Fixed) We have fixed the model to Gemma3-12B. While our infrastructure supports multiple models, frequently switching between them incurs significant performance overhead. Fixing the model ensures that every developer gets fast and consistent response times.
  • Context Window: 32K Tokens Your prompts and the generated responses share a context window of 32000 tokens. This is the total memory the model has for a single conversation turn. Be mindful of this limit when designing long, conversational agents.
  • Session Management (KV Cache) A “Session” in our service refers to the model’s short-term memory (the KV-Cache) of your ongoing conversation. Here’s what you need to know:
    • Persistence is Not Guaranteed: We do not promise to retain your session’s KV-Cache indefinitely.
    • System-Wide Limit: The service retains a maximum of 100 active sessions across all users.
    • Developer Takeaway: This means your agent can have short, stateful conversations. However, for long-term memory, you must manage the conversational history within your own AO process and pass relevant context in each new prompt.