Hy3: A High-Efficiency Reasoning and Agent Model

Hy3 is a 295B parameter Mixture-of-Experts (MoE) model, featuring 21B active parameters, designed for advanced reasoning and agentic tasks. It offers significant cost efficiency and performance competitive with much larger models, making it a powerful option for developers and researchers seeking production-ready AI capabilities across diverse applications like coding, office automation, and financial modeling.

How It Works

Hy3 utilizes a Mixture-of-Experts (MoE) architecture, comprising 295 billion total parameters but activating only 21 billion per inference, enhancing efficiency. It incorporates a specialized 3.8 billion parameter MTP (Mixture-of-Turing-Projections) layer and 192 experts with top-8 activation. This design, coupled with extensive post-training and RL scaling, enables strong performance in reasoning, agentic tasks, and handling a 256K context length, outperforming models with significantly more parameters.

Quick Start & Requirements

• Primary deployment via vLLM or SGLang, offering OpenAI-compatible APIs.
• Non-default prerequisites: Python 3.12 (for vLLM build), transformers>=5.6.0 (for SGLang). Serving requires substantial hardware, with 8 GPUs (e.g., H20-3e) recommended for the 295B model.
• Links: vLLM recipes, SGLang cookbook.
• Example commands are provided for starting vLLM and SGLang servers with MTP enabled.

Highlighted Details

• Outperforms similar-sized models and rivals flagship open-source models with 2-5x parameters.
• Achieved 2.67/4 in a blind expert evaluation, surpassing GLM-5.1 (2.51/4), with notable strengths in frontend development, data & storage, and CI/CD.
• Significantly reduced hallucination rate (from 12.5% to 5.4%) and commonsense error rates (from 25.4% to 12.7%) through fine-grained data cleaning and training constraints.
• Improved multi-turn intent tracking and complex context retention, reducing issue rates from 17.4% to 7.9%.

Maintenance & Community

Developed by the Tencent Hy Team, with a contact email provided for R&D and product teams (hunyuan_opensource@tencent.com). No specific community channels like Discord or Slack, nor a public roadmap, are detailed in the provided documentation.

Licensing & Compatibility

Released under the Apache License 2.0. This license is generally permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Serving the full 295B parameter model necessitates substantial hardware resources, with 8 GPUs recommended for optimal performance. While the model demonstrates strong capabilities, specific performance claims are based on internal evaluations and benchmarks, requiring users to validate suitability for their unique use cases.