Summary Hy3 preview is a 295B parameter Mixture-of-Experts (MoE) model from Tencent's Hy Team, excelling in complex reasoning, instruction following, context learning, coding, and agent tasks. Targeting researchers and power users, it offers leading performance with significant cost efficiency due to its 21B active parameters, making advanced AI capabilities more accessible.

How It Works This MoE model features 295 billion total parameters with only 21 billion active during inference, including a specialized 3.8B parameter MTP layer. It comprises 80 standard layers and one MTP layer, utilizing GQA attention and a substantial 256K context length. This architecture enables efficient processing of lengthy contexts and complex instructions, driving significant improvements in reasoning, coding, and agentic behaviors.

Quick Start & Requirements Deployment is recommended via vLLM or SGLang, both requiring building from source. Serving the model necessitates substantial hardware, specifically 8 GPUs with large memory capacity (e.g., H20-3e). Model weights are available on Hugging Face, ModelScope, and GitCode.

• Hugging Face: https://huggingface.co/tencent/Hy3-preview
• ModelScope: https://modelscope.cn/models/Tencent-Hunyuan/Hy3-preview
• GitCode: https://ai.gitcode.com/tencent_hunyuan/Hy3-preview

Highlighted Details

• Advanced Reasoning: Achieves top scores on challenging STEM benchmarks like FrontierScience-Olympiad and IMOAnswerBench, demonstrating strong generalizable reasoning.
• Contextual Understanding: Exhibits solid gains in context learning and instruction following, validated on custom business-scenario benchmarks.
• Agent Capabilities: Shows significant improvements in coding and agent tasks, with competitive performance on SWE-bench, Terminal-Bench 2.0, BrowseComp, and WideSearch.
• Cost Efficiency: Outperforms larger models on key benchmarks while activating fewer parameters (21B vs. 32B-37B for competitors).

Maintenance & Community Developed by the Tencent Hy Team, Hy3 preview was open-sourced on April 23, 2026. Availability on major model hubs facilitates community engagement.

Licensing & Compatibility Hy3 preview is released under the Tencent Hy Community License Agreement. Specific terms regarding commercial use or integration into closed-source projects are not detailed and require consulting the full license text.

Limitations & Caveats As a "preview" release, the model may be subject to future changes. Deployment requires significant GPU resources (8 large-memory GPUs). The "Community License" may impose usage restrictions, necessitating a review of the full license agreement.