Hy3 API Guide: How to Use Tencent Hunyuan Hy3 on SiliconFlow

Tencent Hunyuan Hy3 is now available on SiliconFlow as tencent/Hy3, giving developers direct access to a powerful MoE model through an OpenAI-compatible LLM API. With 295B total parameters, 21B active parameters, and long-context capability, Hy3 is built for tasks that need more than short answers. It fits coding agents, long-document analysis, multi-step reasoning, tool-use workflows, and production AI applications that need strong context handling.

Hy3 is available on SiliconFlow with free API access for the first two weeks after launch, giving developers a practical window to test real workloads, measure API behavior, and decide how Hy3 fits into their application stack.

What Is Tencent Hunyuan Hy3?

Tencent Hunyuan Hy3 is a Mixture-of-Experts model developed by Tencent Hy Team. It uses a sparse MoE architecture, where each token is routed through selected experts instead of activating the full model.

Hy3 has 295B total parameters and 21B active parameters. This architecture supports complex reasoning, instruction following, coding, context learning, and agent tasks while keeping inference focused on the active expert paths.

Hy3 also supports long-context workloads. This is important for AI applications that need to process large inputs, such as:

• Long technical documents

• Multi-file code projects

• Extended customer conversations

• Research notes

• Product specifications

• Legal or policy documents

• Multi-turn agent memory

For developers using SiliconFlow, the model ID is:

"model": "tencent/Hy3"

You can call Hy3 through the same Chat Completions workflow used for other LLM API integrations on SiliconFlow. 

Why Hy3 Matters for Agentic AI and Long-Context Tasks

Many AI applications are moving from simple chatbots to agentic systems. These systems do not only answer questions. They read context, call tools, inspect files, make plans, revise outputs, and continue across multiple steps.

Hy3 fits this shift because agentic workflows often need both reasoning depth and a large working context. A single request may include system instructions, developer rules, tool definitions, user input, retrieved documents, previous steps, and intermediate observations. With long-context support, Hy3 gives developers more room to preserve the full task state instead of compressing every detail into a short prompt.

This is especially useful for coding and automation workflows. A coding agent may need to compare several files, trace a bug across modules, explain a repository structure, generate tests, or refactor code while preserving existing behavior. A general-purpose agent may need to review long documents, call external tools, summarize results, and continue from earlier decisions.

SiliconFlow also makes Hy3 easier to use inside existing agent and coding-tool environments. Because SiliconFlow provides an OpenAI-compatible API, developers can connect SiliconFlow models to tools that support custom OpenAI-compatible providers. This includes agent and coding tools such as Claude Code, Hermes Agent, OpenClaw, OpenCode, Cline, Continue, and similar IDE or agent environments that allow a custom base URL, API key, and model ID.

In this setup, developers do not need to rebuild their workflow from scratch. They can configure the SiliconFlow API endpoint, enter their API key, set the model as tencent/Hy3, and use Hy3 inside the tools they already use for coding or agentic work.

For document workflows, Hy3 can help with report review, contract analysis, policy comparison, long-form summarization, and structured extraction. Instead of treating each paragraph as an isolated input, developers can preserve more source material in one request and ask Hy3 to reason across the full context.

For teams building agentic AI applications, this means Hy3 can be tested in realistic workflows: coding agents, IDE assistants, document agents, internal copilots, and tool-using automation systems. The value is not only the model itself, but also the ability to call it through SiliconFlow’s LLM API across common developer tools and agent frameworks.

How to Call Hy3 With the SiliconFlow API

Hy3 can be called through the SiliconFlow Chat Completions endpoint.

The basic request needs three things:

• Your API key

• The model ID: tencent/Hy3

• A messages array

Here is a simple cURL example:

curl --request POST \
  --url https://api.siliconflow.com/v1/chat/completions \
  --header "Authorization: Bearer YOUR_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "tencent/Hy3",
    "messages": [
      {
        "role": "system",
        "content": "You are a senior software engineer. Give clear and practical answers."
      },
      {
        "role": "user",
        "content": "Review this API error-handling logic and suggest improvements."
      }
    ],
    "temperature": 0.6,
    "max_tokens": 1200
  }'

curl --request POST \
  --url https://api.siliconflow.com/v1/chat/completions \
  --header "Authorization: Bearer YOUR_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "tencent/Hy3",
    "messages": [
      {
        "role": "system",
        "content": "You are a senior software engineer. Give clear and practical answers."
      },
      {
        "role": "user",
        "content": "Review this API error-handling logic and suggest improvements."
      }
    ],
    "temperature": 0.6,
    "max_tokens": 1200
  }'

curl --request POST \
  --url https://api.siliconflow.com/v1/chat/completions \
  --header "Authorization: Bearer YOUR_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "tencent/Hy3",
    "messages": [
      {
        "role": "system",
        "content": "You are a senior software engineer. Give clear and practical answers."
      },
      {
        "role": "user",
        "content": "Review this API error-handling logic and suggest improvements."
      }
    ],
    "temperature": 0.6,
    "max_tokens": 1200
  }'

For Python applications, you can use an OpenAI-compatible client by setting the SiliconFlow base URL:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.siliconflow.com/v1"
)

response = client.chat.completions.create(
    model="tencent/Hy3",
    messages=[
        {
            "role": "system",
            "content": "You are a helpful AI assistant for software development."
        },
        {
            "role": "user",
            "content": "Explain how to redesign this function for better readability."
        }
    ],
    temperature=0.6,
    max_tokens=1200
)

print(response.choices[0].message.content)

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.siliconflow.com/v1"
)

response = client.chat.completions.create(
    model="tencent/Hy3",
    messages=[
        {
            "role": "system",
            "content": "You are a helpful AI assistant for software development."
        },
        {
            "role": "user",
            "content": "Explain how to redesign this function for better readability."
        }
    ],
    temperature=0.6,
    max_tokens=1200
)

print(response.choices[0].message.content)

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.siliconflow.com/v1"
)

response = client.chat.completions.create(
    model="tencent/Hy3",
    messages=[
        {
            "role": "system",
            "content": "You are a helpful AI assistant for software development."
        },
        {
            "role": "user",
            "content": "Explain how to redesign this function for better readability."
        }
    ],
    temperature=0.6,
    max_tokens=1200
)

print(response.choices[0].message.content)

For long responses, set stream to true. Streaming is useful when you are building chat interfaces, coding tools, document assistants, or agent dashboards.

{
  "model": "tencent/Hy3",
  "messages": [
    {
      "role": "user",
      "content": "Summarize this long technical document and extract implementation risks."
    }
  ],
  "stream": true,
  "max_tokens": 1600
}

{
  "model": "tencent/Hy3",
  "messages": [
    {
      "role": "user",
      "content": "Summarize this long technical document and extract implementation risks."
    }
  ],
  "stream": true,
  "max_tokens": 1600
}

{
  "model": "tencent/Hy3",
  "messages": [
    {
      "role": "user",
      "content": "Summarize this long technical document and extract implementation risks."
    }
  ],
  "stream": true,
  "max_tokens": 1600
}

In production, avoid scattering API logic across your codebase. Build a reusable wrapper for Hy3 requests. That wrapper can manage timeout settings, retries, logging, prompt templates, token limits, and fallback logic.

Key Parameters to Check Before Running Hy3

A strong Hy3 integration starts with the right request settings. These parameters affect output quality, latency, cost, and reliability.

Model

Use tencent/Hy3 as the model value.

"model": "tencent/Hy3"

Keep this value consistent across testing and production. If your team runs A/B tests with other models, log the model name with each response.

Messages

The messages array defines the conversation. Use the system role for stable behavior instructions. Use the user role for the actual task. Use previous assistant messages only when the conversation history is needed.

For long-context tasks, organize the input clearly. Use headings, separators, and short instructions. Do not mix business rules, source documents, and output format requirements in one unstructured block.

Temperature

Use a lower temperature for tasks that need stable answers. This includes extraction, code review, classification, compliance checks, and structured summaries.

Use a higher temperature for brainstorming, rewriting, naming, and creative exploration.

Max Tokens

max_tokens controls the response length. Do not set it blindly. Estimate the output you actually need.

For example:

• Short classification: 100–300 tokens

• Code review summary: 800–1500 tokens

• Long document summary: 1500–3000 tokens

• Multi-step agent response: depends on workflow design

Leave enough room between input size and output size. Long-context support does not mean every request should use the full context window.

Streaming

Enable streaming when users need fast visible feedback. This helps with long answers, coding suggestions, and report generation.

For backend-only jobs, non-streaming requests may be easier to log and evaluate.

Tools

For agentic AI, tool use can be important. When function calling is enabled for your workflow, Hy3 can help decide when to call external functions and continue after tool results return.

Keep tool definitions short. Use clear names. Avoid overlapping tools that do similar things. Test whether the model selects the right tool before relying on it in production.

Best Use Cases for Hy3 in Real Applications

Hy3 is most useful when the task needs more context, stronger reasoning, or multi-step execution. It is not limited to basic chatbot replies.

Coding Agents and Developer Tools

Hy3 can support code explanation, bug analysis, pull request review, test generation, frontend development tasks, CI/CD-related analysis, and cross-file refactoring. It is especially useful when the model needs to understand more than one function, file, or intermediate tool result.

Good test cases include:

• Refactoring a service module

• Explaining a repository structure

• Finding edge cases in API logic

• Generating unit tests from existing code

• Reviewing a migration plan

• Checking frontend implementation details

• Analyzing CI/CD failure patterns

Long-Document Analysis

Hy3 can process long inputs such as manuals, contracts, research reports, policies, and product specifications. Developers can build workflows for summarization, comparison, risk extraction, and structured output.

For better results, divide long documents with headings. Tell the model what to prioritize. Ask for a specific output format when needed.

Agentic Workflow Automation

Hy3 can help power agents that plan, retrieve information, call tools, and revise outputs. These workflows are useful for internal productivity tools, customer support assistants, research copilots, and operations automation.

Test the full agent loop. A single good answer does not prove the agent is production-ready.

Enterprise Knowledge Applications

Many enterprise AI tools need to process messy internal information. Hy3 can help with support tickets, meeting notes, product feedback, project plans, and technical knowledge bases.

For this use case, combine Hy3 with retrieval, logging, and human review. Long context is valuable, but clean information flow still matters.

Prompt Evaluation and Model Comparison

Hy3 is also useful for model evaluation. Developers can compare short prompts, long prompts, tool-use prompts, and structured-output prompts under the same LLM API workflow.

How to Use the Free First Two Weeks to Test Hy3

The first two weeks of free Hy3 access should be treated as an engineering evaluation window, not just a trial period. Use it to test real workloads before moving into production.

Start with a small benchmark set. Include the tasks your application actually needs to handle.

A practical test set may include:

• 20 short chat prompts

• 20 long-context prompts

• 10 coding tasks

• 10 document-analysis tasks

• 10 tool-use or agent tasks

• 10 difficult edge cases

For each test, track more than answer quality. Measure latency, token usage, output length, retry rate, and failure patterns.

For coding workflows, test real code instead of toy examples. Include incomplete files, unclear comments, dependency issues, and edge cases. This shows how Hy3 behaves when the input is close to production reality.

For long-document workflows, test different prompt structures. Put the task goal near the top. Separate source content from instructions. Ask for concise outputs when the input is very long.

For agent workflows, test tool selection and recovery. Check whether Hy3 chooses the right function, passes usable arguments, and continues the task after the tool returns a result.

The free period is also a good time to estimate future cost. Record average input tokens, output tokens, retries, and expected daily traffic. This gives your team a clearer production plan.

Move From Hy3 Testing to Production

After testing, move slowly from prototype to production. Do not rely only on a few successful demo prompts.

Start by locking the prompt format. Keep system instructions stable. Define what the model should do when information is missing. Add fallback behavior for timeouts, empty responses, and malformed inputs.

Next, monitor real usage. Track latency, token consumption, user feedback, and failed requests. For agent workflows, log tool calls and intermediate steps. This makes debugging much easier.

Also create evaluation examples. Save strong answers, weak answers, and edge cases. Re-run these examples when you change prompts, tools, or application logic.

Hy3 gives developers a strong foundation for long-context and agentic AI applications. A reliable product still needs careful prompt design, observability, and production testing. 

Start Building With Hy3 on SiliconFlow

Hy3 is now available on SiliconFlow through tencent/Hy3, giving developers a direct path to test Tencent Hunyuan’s MoE model through an OpenAI-compatible LLM API.

Use the free first two weeks to test real documents, real code, real agent flows, and real production-style prompts. Then refine your prompts, set practical token limits, monitor latency, and prepare your application for stable deployment.

FAQs

Q1. What Is the Hy3 API?

The Hy3 API lets developers call Tencent Hunyuan Hy3 through SiliconFlow without deploying the model themselves. Use tencent/Hy3 with the Chat Completions endpoint.

Q2. Is Hy3 a Good Model for Agentic AI?

Yes. Hy3 is well suited for agentic AI because it supports long-context tasks, complex instruction following, coding workflows, and tool-use scenarios. Developers should still test full agent loops before production.

Q3. How Is Hy3 Different From a Standard LLM API Model?

Hy3 uses a Mixture-of-Experts architecture with 295B total parameters and 21B active parameters. It is designed for more demanding workloads such as long-document analysis, code tasks, and multi-step reasoning.

Q4. Can I Use Hy3 for Coding Tasks?

Yes. Hy3 is suitable for coding and agentic development workflows, including code explanation, code review, test generation, refactoring, frontend tasks, CI/CD-related analysis, and repository-level reasoning. It is also designed to work across tool-use and agent scaffolding scenarios, so developers can test it in coding agents, IDE assistants, and multi-step development workflows. For best results, provide enough surrounding code, clear task instructions, and realistic project context.

Q5. How Should I Test Hy3 During the Free First Two Weeks?

Use the free period to test real workloads. Measure answer quality, latency, token usage, long-context behavior, tool-use reliability, and failure cases. This gives your team better data before production use.