Introduction
In this tutorial, you'll learn how to work with Mixture-of-Experts (MoE) models like Tencent's Hy3, which is a 295 billion parameter model that only activates 21 billion parameters per token. This approach allows for efficient processing of long-context tasks while maintaining high performance. We'll walk through how to access and use such models through the OpenRouter platform, which is where you can try Hy3 for free until July 21, 2026.
Prerequisites
- A basic understanding of what large language models (LLMs) are
- An active internet connection
- A web browser (Chrome, Firefox, Safari, etc.)
- No programming experience required - this is a beginner-friendly tutorial
Step-by-Step Instructions
1. Understanding Mixture-of-Experts (MoE) Models
1.1 What is an MoE Model?
Mixture-of-Experts models are a type of neural network architecture that uses multiple smaller models (called "experts") to process information. Instead of using all parameters for every input, only a subset of experts is activated for each token (a word or part of a word). This makes the model more efficient and allows it to handle longer contexts.
1.2 Why Use MoE Models?
MoE models like Hy3 offer several advantages:
- They can process longer text sequences (up to 256K tokens in Hy3's case)
- They use fewer active parameters per token, reducing computational cost
- They maintain high performance while being more scalable
2. Accessing Hy3 via OpenRouter
2.1 Visit the OpenRouter Website
Navigate to https://openrouter.ai in your web browser. This is the platform where you can access Hy3 and other cutting-edge models.
2.2 Create an Account (if needed)
If you don't already have an account, click on "Sign Up" or "Log In" and follow the prompts to create a free account. You'll need to provide an email address and create a password.
2.3 Find Hy3 in the Model Library
Once logged in, look for the search bar or model selection dropdown. Type "Hy3" into the search box. You should see Tencent's Hy3 model listed with its specifications, including the 295B parameters and 21B active parameters.
2.4 Select Hy3 for Use
Click on the Hy3 model card to view its details. Make sure to note the context window size (256K tokens) and the fact that it's free to use until July 21, 2026.
3. Using Hy3 for Your Tasks
3.1 Prepare Your Prompt
Before sending a request to Hy3, you need a clear prompt. For example, if you're doing a coding task, you might write:
Write a Python function that finds the longest common subsequence between two strings.
For a reasoning task, try:
Explain the implications of quantum computing on current encryption methods.
3.2 Send Your Prompt to Hy3
After selecting Hy3, you'll see an interface where you can enter your prompt. Type or paste your question or task into the text box. Hy3 will process your input using its 21 billion active parameters, even though the full model has 295 billion parameters.
3.3 Review the Output
Once you submit your prompt, Hy3 will generate a response. Pay attention to:
- How well the model understands your request
- The quality of the reasoning or explanation provided
- Whether the output shows lower hallucination rates compared to other models
3.4 Experiment with Different Tasks
Try different types of prompts to see how Hy3 performs:
- Code generation tasks
- Long-context summarization
- Multi-step reasoning problems
- Question answering with complex queries
4. Understanding the Context Window
4.1 What is a Context Window?
The context window is the maximum number of tokens (words or subwords) that a model can consider at once. Hy3's 256K context window means it can process up to 256,000 tokens in a single input, which is ideal for:
- Long document analysis
- Complex multi-turn conversations
- Research tasks requiring extensive background information
4.2 How to Use Long Contexts
To take advantage of Hy3's long context capabilities, provide it with substantial input text:
Here is a detailed research paper on machine learning. Summarize the key findings in bullet points.
This allows Hy3 to reference the entire paper rather than just a few sentences.
5. Observing Performance Metrics
5.1 Hallucination Rate
One key advantage of Hy3 is its lower hallucination rates. This means the model is less likely to generate false or misleading information. When you use Hy3, pay attention to whether the output seems factual and consistent with known information.
5.2 SWE-Bench Verified Score
Hy3 achieved a score of 78.0 on SWE-Bench Verified, a benchmark for code generation and reasoning. This score indicates its effectiveness in software-related tasks.
6. Best Practices for Using Hy3
6.1 Start Simple
Begin with straightforward prompts and gradually increase complexity as you become more comfortable with the model's behavior.
6.2 Provide Clear Instructions
When asking for specific tasks, be as clear as possible about what you want. For example:
Please write a Python script that sorts a list of dictionaries by a specific key.
6.3 Test Long-Context Tasks
Use Hy3's long context window for tasks that require extensive background information or multi-step reasoning.
Summary
In this tutorial, you've learned how to access and use Tencent's Hy3 MoE model through the OpenRouter platform. You've discovered how MoE models work by activating only a subset of parameters per token, making them efficient for long-context tasks. You've also learned how to prepare prompts, understand the model's 256K context window, and observe its lower hallucination rates compared to other models. With this knowledge, you can now experiment with Hy3 for various reasoning, agentic, and long-context tasks while enjoying its free access until July 21, 2026.
