Hy3 - Tencent Hunyuan's open-source mixture-of-experts model

What is Hy3

Hy3 is a 295B-parameter mixture-of-experts (MoE) model open-sourced by Tencent Hunyuan. The model has made significant progress in agent capabilities, reasoning, and long-context tasks, with multiple benchmarks rivaling flagship models with 2-5x the parameter scale. Through reinforcement post-training and data cleaning, the hallucination rate has been reduced to 5.4%, and the common-sense error rate to 12.7%. Hy3 supports diverse productivity scenarios such as code development, office automation, and front-end design, making it a cost-effective and reliable choice.

Main features of Hy3

• Agent tasks: Significant progress in tasks such as reasoning, agents, and long context, with multiple benchmarks rivaling flagship models with 2-5x the parameter scale.
• Code development: Supports software development, front-end design, CI/CD, etc., with excellent performance on the SWE-bench series of benchmarks.
• Office productivity: Supports Excel modeling and analysis, PPT creation, complex data aggregation, and linked formula calculations.
• Tool calling: Enhances tool-calling stability and cross-scaffold generalization, supporting Codebuddy, Cline, KiloCode, and more.
• Anti-hallucination: Through fine-grained data cleaning and training constraints, significantly reduces the hallucination rate and common-sense error rate.
• Multi-turn dialogue: Optimizes coreference resolution, ellipsis restoration, and multi-turn constraint inheritance, maintaining intent consistency in complex long-range interactions.

Technical principles of Hy3

• Post-training optimization: Based on preview version to improve the quality and diversity of post-training data and expand the scale of RL compute.
• Joint training: Jointly optimizes business pain points such as coreference resolution, ellipsis restoration, and multi-turn constraint inheritance during the SFT and RL stages.
• Data cleaning: Performs fine-grained data cleaning and training constraints, following the principle of "answer only when there is evidence, and clearly indicate missing evidence when there is none."
• Software-hardware co-optimization: Reduces inference costs and improves Token efficiency through software-hardware co-optimization.

Follow WeChat and reply "open source" to joinAI Open Source Project Communication Group

How to use Hy3

• Obtain model weights: Visit GitHub and HuggingFace to download the model weights and code.
• Activate the API service: Activate the Hy3 API service in the Tencent Cloud console, obtain an access key, and then call it through the interface.
• Configure the runtime environment: Configure the runtime environment locally or in the cloud, load the model, and connect it to Codebuddy、Cline、KiloCode and other scaffold tools.
• Call to complete tasks: Call the model via API or local deployment to complete tasks such as code development, office automation, and data analysis.

Core advantages of Hy3

• Leapfrog performance: Significantly leads in intelligence level among models of the same size, with multiple benchmarks rivaling flagship models with 2-5x the parameter scale.
• Reliable experience: The model's hallucination rate dropped from 12.5% to 5.4%, the common-sense error rate dropped from 25.4% to 12.7%, and the multi-turn problem rate dropped from 17.4% to 7.9%.
• Stable tools: The standard deviation of SWE Bench Verified scores across different scaffolds is controlled within 4 percentage points, and the error recovery capability of tool calling is greatly improved.
• Low cost: API input 1 yuan per million Tokens, output 4 yuan per million Tokens, cached-hit input only 0.25 yuan, and Token consumption for office tasks is reduced by about 47-49% compared with competing products.

Hy3's project address

• Project official website：https://hy.tencent.com/research/hy3
• GitHub repository：https://github.com/Tencent-Hunyuan/Hy3
• HuggingFace model library：https://huggingface.co/tencent/Hy3

Comparison of Hy3 with similar competing products

Hy3's application scenarios

• Software development: Supports code generation, debugging, and CI/CD process automation, with excellent performance on benchmarks such as SWE-bench.
• Office automation: Supports Excel modeling and analysis, PPT creation, complex data summarization, and linked formula calculation.
• Front-end design: Can generate directly runnable front-end code such as HTML web pages and browser extensions.
• Intelligent customer service: Supports AI avatars and customer service scenarios, and can handle incomplete expressions and make reasonable judgments based on context.
• Game development: Supports game logic writing and game assistant integration, such as the AI assistant for Path of Exile: Awakening.