Hy3 preview - Tencent Hunyuan's open-source hybrid expert model integrating fast and slow thinking

What is Hy3 preview

Hy3 preview is Tencent Hunyuan's open-sourcefast and slow thinking integratedmixture-of-expertsmodel, positioned as "the most intelligent model to date". The model adopts a MoE architecture, using 21B activated parameters to achieve a total parameter scale of 295B, balancing performance and inference cost. The model focuses on comprehensive practicality, refusing to be "unbalanced", and emphasizes the systematic synergy of capabilities such as reasoning, long text, instructions, dialogue, code, and tools. The model verifies its real combat power through self-built evaluations, the latest exams, and product public testing, and deeply collaborates with inference framework optimization to increase inference efficiency by 40%, making high-performance AI commercially usable.

Main features of Hy3 preview

• Complex reasoning: Demonstrates generalizable strong reasoning ability in math olympiads, biology leagues, and frontier scientific problems.
• Code and agents: Supports one-time generation of complete mini-program/mini-game code, with outstanding performance in evaluations such as SWE-Bench Verified and Terminal-Bench 2.0.
• In-context learning: Based on the self-developed CL-bench and CL-bench-Life benchmarks, accurately understands implicit constraints and rules in messy long texts.
• Instruction following: Accurately handles complex scheduling, project planning, and other real work scenario needs in multi-turn dialogues.
• Natural dialogue: Reduces the "machine feel", empathizes first before answering, writes more naturally, and uses more vivid metaphors.
• Long text processing: Supports 256K context, and can handle full-text understanding and summarization of documents with tens of thousands of words.

Technical principles of Hy3 preview

• MoE mixture-of-experts architecture: Adopts a sparse activation mechanism, activating only 21B parameters per inference, achieving a balance between high performance and low inference cost at a total parameter scale of 295B.
• Fast and slow thinking fusion mechanism: Integrates two cognitive modes, fast intuitive reasoning and slow deep thinking, dynamically allocating computational resources and reasoning paths according to task complexity.
• Pre-training infrastructure reconstruction: Tears down and rebuilds the pre-training and reinforcement learning framework, expands the scale of RL training, and specifically improves code generation and agent task execution capabilities.
• Architecture-inference collaborative optimization: Deeply collaborates model structure design with the underlying inference framework, and through engineering methods such as operator fusion and memory optimization, increases end-to-end inference efficiency by 40%.
• Long context processing mechanism: Supports 256K tokens ultra-long context, adopts efficient position encoding and sparse attention mechanisms, and achieves precise full-text localization and understanding of documents with tens of thousands of words.
• Systematic multi-capability training: Refuses single-capability "imbalance", and through a unified training framework achieves deep synergy and mutual enhancement of capabilities such as reasoning, code, dialogue, instruction following, and tool calling.

How to use Hy3 preview

• Direct experience on the official website: Visit theTencent Hunyuanofficial website to chat with the model online and test reasoning, code generation, and long text understanding capabilities.
• Open-source local deployment: Go to GitHub or Hugging Face and search for "Tencent Hy3 preview", download the model weights and inference code, and deploy and fine-tune based on a local GPU environment.
• API calling development: Log in to Tencent Cloud TokenHub, select a Hy3 preview plan (Lite/Standard/Pro/Max), and after obtaining an API Key, you can integrate the model's capabilities into your own applications or workflows through standard interfaces.
• Direct use in Tencent products: InYuanbao、ima、CodeBuddy、WorkBuddy, QQ, QQ Browser, Tencent Docs, Tencent Lexiang, Sogou Input Method, Tencent Maps, Tencent E-Sign, Tencent Cloud and other products, Hy3 preview has been fully launched as the underlying model, and the new model capabilities can be invoked directly through conversation.

Key information and usage requirements of Hy3 preview

• Model architecture: Fast-slow thinking fusion MoE, with 295B total parameters and 21B activated parameters.
• Context length: Supports up to 256K tokens.
• Open source status: Model weights and code are fully open source on GitHub and Hugging Face, free to download.
• API calls: Can be called through Tencent Cloud TokenHub, with input starting at 1.2 yuan per million tokens and output starting at 4 yuan per million tokens.
• Plan pricing: Lite plan 28 yuan/month (35 million tokens), Standard 78 yuan/month, Pro 238 yuan/month, Max 468 yuan/month.
• Integrated products：Yuanbao、ima、CodeBuddy、WorkBuddy, QQ, QQ Browser, Tencent Docs, Tencent Lexiang, Sogou Input Method, Tencent Maps, Tencent E-Sign, Tencent Cloud, etc.

Core advantages of Hy3 preview

• Real-evaluation oriented: Moving beyond public leaderboards that are easy to game, capabilities are verified through 50+ self-built evaluation suites, real exams, and product crowdsourced testing.
• High cost-effectiveness: Deep co-design of model architecture and inference framework, improving inference efficiency by 40%, with only 21B activated parameters.
• Leap in Agent capabilities: Significant progress in code generation and multi-step task execution, capable of outputting a complete runnable project in one go.
• Pragmatism principle: The model pursues the unity of systematic capabilities, authentic evaluation, and commercial reasonableness.

Project addresses of Hy3 preview

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

Comparison of Hy3 preview with similar competitors

Application scenarios of Hy3 preview

• Academic Research: Derive complex mathematical formulas, solve Olympiad-level science problems, assist with paper reading and scientific hypothesis validation.
• Educational Tutoring: Explain highly difficult problems, dynamically adjust the depth of explanation according to the student's level, and generate personalized practice problems and knowledge summaries.
• Software Development: Output complete WeChat Mini Program/Mini Game code and configuration files in one go, automatically fix code bugs, and assist with front-end and back-end joint debugging and terminal task execution.
• Smart Office: Accurately extract to-dos and schedules from messy meeting minutes, handle complex cross-department project coordination, and directly generate PPT content in Tencent Docs.
• Content Operations: Write WeChat Official Account articles and marketing copy, optimize text to reduce the "AI tone", and generate creative content and vivid metaphors in multiple styles.