Tencent Hy3 Model Overview

Executive Summary

Tencent has released the full version of Hy3, a mixture-of-experts (MoE) model designed to compete in the mid-tier frontier model space. Hy3 is specifically optimized for agentic tasks, tool use, and local deployment, positioning itself as a more hardware-efficient alternative to larger models like GLM 5.2 while maintaining high performance in reasoning and hallucination reduction.

Model Architecture and Specifications

Hy3 is built as a large-scale mixture-of-experts model with a focus on balancing power and efficiency.

• Parameter Count: 295 billion total parameters.
• Active Parameters: 21 billion active parameters per token.
• Speculative Decoding: Includes a 3.8 billion parameter speculative decoding model to increase inference speed.
• Context Window: 256K tokens.

Performance and Benchmarks

Hy3 is positioned as a "mid-tier" model, targeting the space between small local models and massive proprietary frontier models.

Agentic Tasks and Tool Use

Hy3 excels in agentic workflows, specifically tool calling and output formatting. In testing, the model demonstrated high proficiency in:

• Repeated Tool Calls: Successfully handling multiple sequential tool calls.
• Pagination: Managing long-running pagination across twelve different tools.
• Error Recovery: Demonstrating resilience by attempting retries when tool calls return errors rather than giving up.
• Noise Filtering: Identifying relevant information from API responses without being distracted by irrelevant data.

Comparison with GLM 5.2

While Hy3 is powerful, it is not intended to replace all high-end models. Specifically, GLM 5.2 generally outperforms Hy3 in agentic coding tasks. However, Hy3 is significantly smaller (well over half the size of GLM 5.2), making it more viable for local hosting and fine-tuning on private hardware without requiring massive clusters of B200 GPUs.

Reliability and Hallucinations

Tencent has focused heavily on post-training and data cleaning to improve reliability. Compared to its preview version, the full Hy3 model has halved both common sense error rates and hallucination rates.

Capabilities and Testing Results

Reasoning and Chain-of-Thought

Hy3 utilizes a long chain-of-thought (CoT) process. In logic puzzle tests, the model generated significant "thinking" tokens to verify its steps before arriving at a solution. The quality of this internal reasoning is noted as being high, potentially exceeding some other open-source models.

Creative and Technical Generation

• SVG Generation: The model can generate complex SVG code, such as a detailed pelican on a bicycle, showing significant improvement over the preview version.
• HTML/CSS: Hy3 is capable of producing polished, functional website layouts, including opt-in forms and integrated images.
• Long-form Content: In a 5,000-word essay test, the model produced a structured outline (acting as an agentic planning step) and generated approximately 2,500 to 3,000 words, noting its own constraints regarding single-block generation.

Deployment and Accessibility

Hy3 is currently available for testing via OpenRouter. Because of its size and architecture, it is viewed as a strong candidate for companies wanting a fully locked-down, local model that can be fine-tuned for specific corporate use cases on manageable hardware.