Hy3 Now Live on SiliconFlow: Advancing Agent Capabilities, Ready for Real-World Workflows

TL;DR

• Hy3: Tencent Hunyuan's next-generation flagship, is now live on SiliconFlow. 295B total parameters, 21B active per token, 256K context.

• Why it matters: It rivals open-source flagships with 2–5x its parameter count on coding, agentic, and reasoning benchmarks — and scored higher than GLM-5.1 in a blind expert evaluation on real-world workflows.

• Free for first 2 weeks: Try Hy3 on SiliconFlow at no cost for the first two weeks.

• Get started: SiliconFlow provides OpenAI- and Anthropic-compatible APIs for Hy3, so it drops straight into Claude Code, Cline, Hermes Agent, OpenCode, and your existing tools in minutes.

When Tencent's Hy team released Hy3 Preview in April, the model quickly proved its value in production — daily token consumption grew twentyfold. Refined through extensive use by developers worldwide and across Tencent's large-scale real-world business scenarios, Hy3 officially launches. It outperforms similar-size models and rivals flagship open-source models with 2-5x parameters. It also shows significant gains in real-world usability across various products and productivity tasks.

Hy3 is now live on SiliconFlow — free for the first two weeks. Our OpenAI- and Anthropic-compatible APIs let you drop it straight into agent frameworks like Claude Code, Hermes Agent, and OpenClaw, or coding assistants like Kilo Code, Cline, and Roo Code.

Here's what you're working with:

Why Hy3 Stands Out

Stronger Agent Capabilities

Building on the Preview version, Hy3 sharpens its edge across reasoning, agentic, and long-context tasks — competitive with open-source flagships at 2–5× its parameter count. Where Hy3 really stands out is real-world productivity work: coding, office tasks, financial modeling, frontend design, and game development. If your daily workflow involves any of these tasks, Hy3 is a reliable, cost-effective default you can reach for without second-guessing.

For the full benchmark results and technical details, see Tencent's official Hy3 model page on Hugging Face.

Real Work, Real Test

Public benchmarks only tell part of the story. Tencent ran a blind evaluation with 270 domain experts using tasks from their actual work — the kind you'd run through your agent stack every day:

• debugging a failing pipeline

• scaffolding a frontend component

• writing a migration script

Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4, with the advantage most pronounced in frontend development, data & storage, and CI/CD workflows.

More Reliable Product Experiences

• Tool-call stability: Hy3 fixes a batch of reliability issues around tool configuration and output formatting. On SWE-Bench Verified, accuracy stays within 4 points across different agent frameworks — CodeBuddy, Cline, and KiloCode. In practice, that means you can switch from Cline to KiloCode without seeing wild swings.

• Anti-hallucination: Hallucination rate dropped from 12.5% to 5.4%; commonsense error rate from 25.4% to 12.7%. The model now adheres to "answer when grounded, state when evidence is missing, do not conflate sources or fabricate data.", which is particularly important for agents that act on retrieved context — a fabricated endpoint or wrong file path can derail an entire workflow.

• Multi-turn context retention: Issue rate in multi-turn conversations dropped from 17.4% to 7.9%; MRCR long-dialogue scores jumped from 42.9% to 75.1%. Complex intents no longer decay over long-horizon interactions. If you've watched an agent forget a constraint from turn 3 to turn 12, this is the fix.

See Hy3 in Action

We wanted to see how Hy3 would handle a real frontend task — not just generate code, but plan, iterate, and refine like an experienced developer.

Instead of jumping straight into implementation, Hy3 first proposed three different architectural approaches, explaining the trade-offs between visual fidelity, performance, and implementation complexity. After we picked one, it moved through its own plan → build → debug → refine workflow before delivering the final result.

Then gave it one follow-up request: add shimmering water-like reflections that respond to mouse movement. After one more iteration, the interaction was integrated seamlessly into the project.

In total, two rounds of interaction without any manual code edits. The result is the demo below:

Run Hy3 with Your Familiar Tools

Benchmarks and backstory can only tell you so much — the fastest way to know if Hy3 fits your workflow is to run it yourself. And that only takes minutes.

Try it first, no setup

Open the SiliconFlow Playground and chat with Hy3 right in your browser.

Tune temperature and top-p, or run it side by side against Hy3-Preview or DeepSeek-V4-Pro on the same coding prompt to see how it performs.

Plug into your tools

SiliconFlow APIs are both OpenAI- and Anthropic-compatible, so Hy3 drops straight into any tool that supports a custom provider — same SDK, same code, just a new base URL and model string. Live in minutes.

What you need to connect:

Get Started Immediately

1. Try it for free: Start using Hy3 on SiliconFlow, free for the first 2 weeks.

2. Integrate: Use our OpenAI-compatible API. Explore the full API specifications in the SiliconFlow API documentation.

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "tencent/Hy3",
    "stream": False,
    "messages": [
        {
            "role": "user",
            "content": "How many r'\''s are in the word '\''strawberry'\''?"
        }
    ],
    "max_tokens": 4096,
    "enable_thinking": False,
    "thinking_budget": 4096,
    "temperature": 0.7
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.request("POST", url, json=payload, headers=headers)

print(response.text)

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "tencent/Hy3",
    "stream": False,
    "messages": [
        {
            "role": "user",
            "content": "How many r'\''s are in the word '\''strawberry'\''?"
        }
    ],
    "max_tokens": 4096,
    "enable_thinking": False,
    "thinking_budget": 4096,
    "temperature": 0.7
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.request("POST", url, json=payload, headers=headers)

print(response.text)

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "tencent/Hy3",
    "stream": False,
    "messages": [
        {
            "role": "user",
            "content": "How many r'\''s are in the word '\''strawberry'\''?"
        }
    ],
    "max_tokens": 4096,
    "enable_thinking": False,
    "thinking_budget": 4096,
    "temperature": 0.7
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.request("POST", url, json=payload, headers=headers)

print(response.text)

• Business or Sales Inquiries →

• Join our Discord community now →

• Follow us on X for the latest updates →

• Explore all available models on SiliconFlow →