The 295B AI Model That Runs Like a 20B: Hy3 Changes Everything

What Is Hy3?

Hy3 is Tencent's newest model from the Hunyuan team. It uses a Mixture-of-Experts architecture with 295 billion total parameters and 192 experts, with the top 8 activated per token. This keeps 21B parameters active during inference, optimizing the compute-to-intelligence ratio.

The model supports a 256K-token context window and includes a 3.8B multi-token prediction layer to accelerate decoding. Tencent released Hy3 under the Apache 2.0 license, allowing for commercial deployment, managed hosting, and custom fine-tuning.

Following the preview in April 2026, the full GA release arrived on July 6, 2026, incorporating expanded post-training data, scaled reinforcement learning, and feedback from over 50 Tencent product teams.

Key specs at a glance:

Architecture Details

Hy3 uses 80 transformer layers, 64 attention heads, grouped query attention with 8 KV heads, a hidden size of 4096, and an intermediate size of 13,312.

The MoE design aligns the active compute requirements with a 20B-class model, while the larger parameter count supports the reasoning capacity required for complex workflows. Tencent also provides an FP8 version alongside the BF16 weights, offering deployment teams an alternative for managing memory efficiency and throughput on multi-GPU serving stacks.

Preview vs. GA Updates

The April preview was the first release from Tencent's rebuilt training infrastructure. The GA release targets stability, benchmark performance, and agent reliability.

Measurable improvements in the GA release include:

In a blind evaluation with 270 domain experts reported by Tencent, Hy3 scored 2.67 out of 4, compared to GLM-5.1 at 2.51. The primary performance gains were recorded in frontend development, data and storage tasks, and CI/CD operations.

Benchmark Performance

Hy3 was evaluated across coding, reasoning, and agent workloads. The reported scores for the GA release are:

For agent systems, WildClawBench and SkillsBench measure multi-step task execution and open-ended behavior, providing metrics more relevant to production workflows than single-turn evaluations.

Usage and Adoption

Within two weeks of the preview launch, Hy3 processed 3.66 trillion tokens on OpenRouter, a 298% week-over-week increase. The highest token consumption came from applications including Hermes Agent, Claude Code, Kilo Code, OpenClaw, and Cline. This usage profile indicates adoption primarily in coding agents and task execution platforms.

Reported developer use cases highlight:

• Instruction following in long-context workflows

• Configurable reasoning depth controlled via API parameters

• Consistent tool calling across different agent frameworks

• Repository analysis, code generation, and structured extraction

Tencent reported a 20x increase in daily token consumption following the preview release, and a 6x growth in users selecting Hy3 within their internal WorkBuddy tool.

Frontend Generation and Cost Efficiency

In a recent evaluation published on X, the GMI Cloud team tested Hy3 against GLM-5.2, GLM-5.1, and DeepSeek V4 to measure frontend code generation. Using three distinct webpage design prompts, Hy3 outperformed both GLM-5.1 and DeepSeek V4 in visual output quality.

Beyond visual performance, the inference economics strongly favor Hy3. In the same test, Hy3 operated at roughly half the cost of the GLM 5.x series, matching the highly efficient pricing profile of DeepSeek V4. For teams building UI generators or automating frontend design workflows, this positions Hy3 as an open-source model that delivers premium visual design logic without a premium cost.

Performance Metrics

Artificial Analysis measured GMI Cloud at 171.4 tokens per second output throughput for the Hy3 preview, compared to 175.5 tokens per second on SiliconFlow.

GMI Cloud supports function calling and structured JSON output for Hy3, which are required features for integrating model outputs into external tools and APIs.

Running Hy3 on GMI Cloud

GMI Cloud is a benchmarked inference provider for Hy3. The API is OpenAI-compatible, requiring only an update to the base URL and API key.

Basic Chat Completion

import openai

client = openai.OpenAI(
api_key="YOUR_GMI_API_KEY",
base_url="https://api.gmi-serving.com/v1"
)

response = client.chat.completions.create(
model="tencent/Hy3",
messages=[
{"role": "user", "content": "Explain how MoE routing works in transformer models."}
],
)

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

Controlling Reasoning Depth

Hy3 provides three reasoning modes. Use no_think for standard tasks and high for coding or logic workloads.

response = client.chat.completions.create(
model="tencent/Hy3",
messages=[
{"role": "user", "content": "Solve this integral step by step."}
],
extra_body={
"reasoning": {"enabled": True, "effort": "high"}
}
)

To disable reasoning, simply omit the reasoning field:

messages = [
{"role": "user", "content": "Summarize this document in 3 bullets."}
]

response = client.chat.completions.create(
model="tencent/Hy3",
messages=messages,
)

Long-Context Repository Analysis

import openai

client = openai.OpenAI(
api_key="YOUR_GMI_API_KEY",
base_url="https://api.gmi-serving.com/v1"
)

with open("repo_dump.txt", "r") as f:
codebase = f.read()

response = client.chat.completions.create(
model="tencent/Hy3",
messages=[
{
"role": "user",
"content": f"Here is the full repository:\n\n{codebase}\n\nFind all SQL injection vulnerabilities and suggest fixes."
}
],
max_tokens=4096,
)

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

The 256K context window allows Hy3 to process approximately 200,000 words of code or documentation per request, supporting full repository reviews without chunking strategies.

Self-Hosting Hy3 on GMI GPU Clusters

Teams requiring dedicated infrastructure can self-host Hy3 on GMI Cloud's GPU clusters. Tencent provides deployment configurations for both vLLM and SGLang.

Recommended deployment specifications:

• Hardware: 8x H100 or H200 GPUs

• Serving Framework: vLLM or SGLang

• Precision: BF16 for standard deployment, FP8 for high-throughput tuning

With 21B active parameters, the hardware requirements for serving Hy3 are lower than dense models of similar total size. For GPU availability and configuration, visit console.gmicloud.ai.

Deployment Links

• Serverless API: console.gmicloud.ai

• Model catalog: gmicloud.ai/en/models

• GPU infrastructure: gmicloud.ai/en/gpus

• Model weights: huggingface.co/tencent/Hy3

• API documentation: docs.gmicloud.ai

• Community support: discord.gg/mbYhCJSbF6