Tencent-Hunyuan/Hy3

What it solves

Hy3 is designed to provide a high-performance, cost-effective alternative to massive flagship models. It specifically addresses common LLM failures such as hallucinations, unstable tool-calling formats, and the loss of intent or context during long, multi-turn conversations.

How it works

Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters. It utilizes a Multi-Token Prediction (MTP) layer to improve efficiency and performance. The model was developed through scaled-up post-training with high-quality data and joint optimization of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to enhance reasoning and reliability.

Who it’s for

It is intended for developers and product teams building productivity applications in areas such as coding, frontend design, financial modeling, office work, and game development, as well as those needing a model capable of stable tool-calling and long-context retention.

Highlights

• Efficient Architecture: Uses an MoE design with 21B active parameters and a 256K context length.
• Production-Grade Reliability: Reduced hallucination rates (from 12.5% to 5.4%) and improved tool-call stability.
• Agentic Capabilities: Strong performance in reasoning and agentic tasks, particularly in CI/CD, data storage, and frontend development.
• Flexible Deployment: Supports deployment via vLLM and SGLang with OpenAI-compatible APIs.
• Reasoning Modes: Offers adjustable reasoning effort (no_think, low, high) for varying task complexity.