Hy3 preview - Tencent's hybrid expert model that integrates fast and slow thinking in open source.

What is Hy3 preview?

Hy3 preview is an open-source version of Tencent's Hunyuan platform.The integration of fast and slow thinkingHybrid ExpertModelPositioned as "the most intelligent model to date," the model employs the MoE architecture, achieving a total parameter scale of 295B with only 21B activation parameters, balancing performance and inference cost. The model prioritizes comprehensive practicality, rejecting any "specialization," and emphasizing the systematic collaboration of capabilities such as reasoning, long-text analysis, commands, dialogue, code, and tools. The model's true capabilities are validated through self-built evaluation systems, the latest examinations, and product crowdsourcing. Deep collaborative inference framework optimization improves inference efficiency by 40%, making high-performance AI commercially viable.

Main functions of Hy3 preview

• Complex ReasoningDemonstrates strong, generalizable reasoning ability in math Olympiads, biology competitions, and cutting-edge scientific challenges.
• Code and Intelligent AgentsIt supports generating complete mini-program/mini-game code in one go, and performs outstandingly in evaluations such as SWE-Bench Verified and Terminal-Bench 2.0.
• Contextual learningBased on the self-developed CL-bench and CL-bench-Life benchmarks, it accurately understands the implicit constraints and rules in messy long texts.
• Instructions followedAccurately handle the complex scheduling and project planning needs in real-world work scenarios involving multiple rounds of dialogue.
• Natural DialogueReduce the "machine-like" feel; empathize first and then answer; this will make your writing more natural and your metaphors more vivid.
• Long text processingIt supports 256K context and can handle full-text understanding and summarization of documents with tens of thousands of words.

Hy3 preview's technical principles

• MoE Hybrid Expert ArchitectureIt adopts a sparse activation mechanism, which activates only 21 parameters in a single inference, achieving a balance between high performance and low inference cost with a total parameter scale of 295B.
• Fast and slow thinking integration mechanismIt integrates two cognitive modes: rapid intuitive reasoning and slow, deep thinking, and dynamically allocates computing resources and reasoning paths according to task complexity.
• Pre-training infrastructure reconstruction: Overhaul and rebuild the pre-training and reinforcement learning framework, expand the scale of RL training, and specifically improve code generation and agent task execution capabilities.
• Architecture-Inference Collaborative OptimizationThe deep collaborative model structure design and underlying inference framework improve end-to-end inference efficiency by 40% through engineering methods such as operator fusion and memory optimization.
• Long context processing mechanismIt supports ultra-long contexts of up to 256K tokens and employs efficient positional encoding and sparse attention mechanisms to achieve accurate full-text positioning and understanding of documents containing tens of thousands of words.
• Multi-ability systematic trainingIt rejects the "one-sided" approach of focusing on a single skill and achieves deep synergy and mutual enhancement of abilities such as reasoning, coding, dialogue, instruction following, and tool usage through a unified training framework.

How to use Hy3 preview

• Experience it directly on the official websiteVisit the Tencent Hunyuan official website to interact with the model online and test its reasoning, code generation, and long text understanding capabilities.
• Open source local deploymentGo to GitHub or Hugging Face and search for "Tencent Hy3 preview" to download the model weights and inference code, and deploy and fine-tune them based on your local GPU environment.
• API call developmentLog in to Tencent Cloud TokenHub, select the Hy3 preview package (Lite/Standard/Pro/Max), and obtain the API Key. Then you can integrate the model capabilities into your own applications or workflows through the standard interface.
• Tencent products can be used directly.Hy3 preview has been fully launched as the underlying model in products such as Yuanbao, IMA, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, Tencent Lexiang, Sogou Input Method, Tencent Maps, Tencent Electronic Signature, and Tencent Cloud, allowing users to directly access the new model's capabilities through dialogue.

Key information and usage requirements for Hy3 preview

• Model ArchitectureFast and slow thinking fusion MoE, total parameters 295B, activation parameters 21B.
• Context lengthMaximum support is 256K tokens.
• Open source situationThe model weights and code have been fully open-sourced on GitHub and Hugging Face and are available for free download.
• API callsIt can be called through Tencent Cloud TokenHub, with a minimum input of 1.2 yuan per million tokens and a minimum output of 4 yuan per million tokens.
• Package PriceLite plan: 28 yuan/month (35 million tokens), Standard plan: 78 yuan/month, Pro plan: 238 yuan/month, Max plan: 468 yuan/month.
• Products already integrated: Yuanbao, IMA, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, Tencent Enjoy, Sogou Input Method, Tencent Maps, Tencent Electronic Signature, Tencent Cloud, etc.

Hy3 preview's core advantages

• Realistic evaluation orientation: Break away from publicly available ranking lists that are easily manipulated, and verify our capabilities through 50+ self-built evaluation systems, real exams, and product crowdsourcing testing.
• High cost performanceThe design of a deep collaborative model architecture and inference framework improves inference efficiency by 40% with only 21B activation parameters.
• Agent's ability leapSignificant improvements have been made in code generation and multi-step task execution capabilities, enabling the output of complete, runnable projects in a single step.
• pragmatism principleThe model aims to achieve a balance between systematized capabilities, authentic evaluation, and commercial viability.

Hy3 preview project address

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

Hy3 preview's comparison with similar competing products

Application scenarios of Hy3 preview

• Academic researchIt can derive complex mathematical formulas, solve Olympiad-level science problems, and assist in reading papers and verifying research hypotheses.
• Educational guidanceIt provides explanations of challenging problems, dynamically adjusts the depth of explanation based on students' levels, and generates personalized practice questions and knowledge summaries.
• Software developmentIt can output complete WeChat mini-program/mini-game code and configuration files in one go, automatically fix code bugs, and assist in front-end and back-end integration and terminal task execution.
• Smart OfficeIt can accurately extract to-do items and schedules from messy meeting minutes, handle complex cross-departmental project coordination, and directly generate PPT content in Tencent Docs.
• Content OperationsWrite articles and marketing copy for WeChat official accounts, optimize text to reduce "AI-like" tone, and generate creative content and vivid metaphors in various styles.