Overview
Tencent Hunyuan released and open-sourced its next-generation flagship model Hy4 preview on August 28, 2026, under the Apache 2.0 license, positioned as 'built for productivity.' The model has a total of 770B parameters (approximately 780B including speculative decoding layers), 49B active parameters, supports a 1M context, and is deeply optimized for software engineering, office analytics, game development, and scientific research scenarios. As a core achievement of Tencent's AI strategy, Hy4 preview is led by Chief AI Scientist Yao Shunyu, continuing an iteration cadence of about one major version every two months, and is simultaneously launched in products such as WorkBuddy, CodeBuddy, Yuanbao, and ima. Users can access the API through Tencent Cloud TokenHub and OpenRouter.
Key Features
- Ultra-large-scale MoE architecture: Total parameters of 770B, active parameters of 49B, designed with a 78-layer MoE, each layer containing 256 routed experts and 1 shared expert, activating the top-8 routed experts per token, combined with 4 residual streams and Gated DSA attention, enabling efficient inference and long-context processing.
- 1M ultra-long context: Supports a context window of 1 million tokens, a significant expansion from the previous generation Hy3 preview's 256K, capable of handling long documents, cross-file collaboration, and complex scientific research tasks, paired with IndexCache cross-layer sparse index reuse technology to improve long-sequence processing efficiency.
- Self-improvement closed loop: For the first time, the model participates in the automatic optimization of its own training methods, data strategies, evaluation systems, and underlying operators. The model can autonomously analyze inference bottlenecks and perform operator fusion and communication optimization, improving end-to-end throughput by 31.8% compared to the baseline.
- Production-oriented training: Training data is co-built with high quality by top experts from Tencent's internal software engineering, gaming, finance, and security fields, and is deeply co-designed with products like WorkBuddy, ensuring the model's practicality and reliability in real production scenarios.
- Enhanced multi-scenario capabilities: Software engineering enhances understanding and debugging of long-term development tasks; office analytics improves complex environment understanding and financial analysis; game development supports generating playable prototypes from a single sentence requirement; scientific research has made significant progress in AI research, molecular dynamics, and other fields.
- Open-source and open ecosystem: Open-sourced under the Apache 2.0 license, providing complete model weights and technical documentation, supporting access through Tencent Cloud TokenHub, OpenRouter, and other channels, and simultaneously offering limited-time free trials in multiple Tencent products.
Use Cases
- Software engineering: Supports understanding, planning, debugging, and validation of long-term development tasks, improving visual aesthetics and interaction quality in front-end development, suitable for code generation and refactoring in complex projects.
- Office analytics: Handles data analysis and cross-file collaboration, completing the full workflow from information extraction to delivery of documents, spreadsheets, and presentations, suitable for financial reports and business intelligence scenarios.
- Game development: Directly generates playable prototypes from a single sentence requirement and is proficient in using game engines, accelerating game design and prototype validation processes.
- Scientific research: Provides assistance in AI research, molecular dynamics simulations, condensed matter physics, basic mathematics, and other fields, supporting long-context paper reading and experimental design.
- Agent applications: Achieves automated task execution through products like WorkBuddy and CodeBuddy, handling complex multi-step workflows with a 1M context.
Pros
- Ultra-large parameter scale and active parameter design deliver strong performance on complex tasks, with an internal blind test average score of 2.99/4 from 163 experts, leading compared models in the same evaluation.
- 1M context window supports ultra-long documents and multi-file collaboration, suitable for long-text-intensive scenarios such as research, law, and finance.
- Self-improvement closed-loop mechanism automatically optimizes training and inference efficiency, increasing end-to-end throughput by 31.8% and reducing long-term usage costs.
- Friendly open-source license: Apache 2.0 allows commercial use and secondary development, with convenient access via Tencent Cloud API.
- Deep integration with Tencent products, with simultaneous launches in WorkBuddy, Yuanbao, and ima, allowing users to experience it quickly.
- Pricing continues the inclusive strategy, approximately 6 RMB per million input tokens, offering outstanding cost-effectiveness.
Pricing
Input costs approximately 6 RMB per million tokens; specific output pricing and detailed billing rules are subject to the official Tencent Cloud website.
Summary
Tencent Hunyuan Hy4 preview, as a new-generation open-source flagship, sets a new benchmark for productivity models with 770B total parameters, 49B active parameters, and a 1M context. Its specialized optimizations in software engineering, office analytics, game development, and scientific research, combined with efficiency gains from the self-improvement closed loop, demonstrate strong practical value. Internal blind test results lead competitors in the same evaluation, and the open-source license and inclusive pricing further lower adoption barriers. As an early preview version, the official statement indicates that pre-training and post-training still have room for improvement, and future versions are worth anticipating.
Version History
- Tencent Hunyuan Hy4 preview released and open-sourced (2026-08-28): Tencent Hunyuan released and open-sourced Hy4 preview (Apache 2.0), its next-generation flagship LLM: 770B total params, 49B active, 1M-token context, built for productivity across software engineering, office analytics, game development and research. In an internal blind evaluation by 163 experts on 203 engineering tasks it scored 2.99/4 (GLM-5.3: 2.92, Kimi K3: 2.94). It is the first Hunyuan model to join the loop of optimizing its own training methods, data strategy, evaluation systems and operators (+31.8% end-to-end throughput). Launched day-one in WorkBuddy/CodeBuddy/Yuanbao/ima, with API access via Tencent Cloud TokenHub and OpenRouter.