Today, we are releasing Hy4 preview: 770B total parameters, 49B activated parameters, and a 1M context length, demonstrating outstanding capabilities on real-world productivity tasks in coding, office work, science, and more.
A New-Generation Flagship Model
Hy4 preview has been significantly scaled up in model size, context length, and data volume. Joint advances in pre-training and post-training have brought another major leap in intelligence, securing its place in the top tier of open-source models.
Built for Productivity
Through high-quality data co-development with top experts within Tencent across software engineering, gaming, finance, security, and other fields, Hy4 preview has made significant progress on a wide range of real-world productivity tasks:
• Software engineering: enhanced understanding, planning, debugging, and verification for long-horizon development tasks, with further improvements in visual aesthetics and interaction quality for front-end development;
• Office analytics: significantly improved understanding of complex office environments and financial analysis capabilities, with a focus on optimizing data analysis and cross-file collaboration, covering the complete workflow from information processing to the delivery of documents, spreadsheets, and presentations;
• Game development: enhanced ability to generate a playable prototype directly from a single-sentence prompt, with proficient use of game engines, allowing developers to continuously refine complex game projects through multi-turn interactions;
• Scientific research: significantly improved understanding, reasoning, and problem-solving for complex research problems, with substantial progress across scenarios including AI R&D, molecular dynamics simulation, condensed matter physics, and fundamental mathematics.
Meanwhile, Hy4 preview continues to integrate deeply with products such as CodeBuddy and WorkBuddy, optimizing the real user experience in productivity scenarios. We organized a blind test in which 163 in-house experts evaluated 203 engineering tasks. The results show that Hy4 preview (average score 2.99/4.00) slightly outperformed GLM-5.3 (average score 2.92/4.00; 46.8% win / 12.8% tie / 40.4% loss) and Kimi K3 (average score 2.94/4.00; 51.2% win / 7.9% tie / 40.9% loss).
Building a Shooter Game
By connecting to the Unreal Engine 5 via MCP, it built a shooter game demo from scratch purely through conversation, including environment art setup, gameplay construction, and level design.
Building a Third-Person Adventure Game
In the Unity (U3D) engine, it conversationally built a third-person penguin adventure game demo from scratch, including scene and environment art setup, movement/combat and quest-tracking gameplay construction, and a three-scene flow covering main menu, loading, and the level.
Ancient Architecture Digital Twin Website
Using Three.js with parametric procedural modeling, Canvas-generated IBL and textures, and per-model camera framing, it produced a desktop-first, full-viewport, interactive digital archive website for ancient architecture.
"Here Comes the Goose" Horizontal Interactive Comic Webpage
It generated native WebGL for a horizontal interactive comic page where scrolling drives the camera along the X-axis. Hy4 preview maintained consistent character art style and narrative continuity across multiple chapters, while accurately orchestrating panel composition, camera pacing, and mobile performance.
Corporate Finance Expense Audit in an Office Scenario
The model automatically determined whether invoices were compliant across 72 files, identified the currently effective rules from 3 sets of regulations, and rigorously vetted whether invoice reimbursement requests were compliant.
Financial World Sandbox
It generated a sandbox demonstrating the model's capabilities in financial research, industry-chain understanding, causal reasoning, and valuation analysis. Users can visually observe how a macro event propagates to industry supply and demand, the industrial chain, key companies, and value distribution across the industry, with clearly traceable transmission paths.
Exploring Scientific Research Scenarios
Autonomously Identifying Bottlenecks and Continuously Optimizing Inference
Hy4 preview autonomously analyzed bottlenecks in the inference system and carried out multiple rounds of optimization targeting operator fusion and communication optimization. End-to-end throughput improved by 31.8% over the baseline, with stable gains across different context lengths and concurrency levels—validating its ability to autonomously locate bottlenecks and continuously optimize inference infrastructure.
Small-Model Post-Training Task
Hy4 preview can manage multiple Codex sessions like a researcher to run experiments and continuously adjust exploration directions based on results. In a small-model post-training task, Hy4 preview acted as a researcher coordinating Codex to optimize multiple evaluation targets simultaneously, outperforming Codex's independent exploration on all 8 evaluations—demonstrating its ability to judge direction, organize experiments, and iterate continuously in complex R&D tasks.
Machine-Learned Force Field Molecular Dynamics Simulation
Working with Hyra, Hy4 preview made progress in machine-learned force field molecular dynamics (MD) simulation: on a highly optimized JAX implementation, the 32,512-atom phospholipid bilayer SO3LR system was further accelerated by 2.0×, reaching 54.9 ms/step, and a single high-end GPU can accommodate 300,000 atoms. This opens up greater computational headroom for new materials screening, drug molecule research, and simulation of complex biological systems.
Low-Temperature Quantum Transport Device Design Problem
For a low-temperature quantum transport device design problem in condensed matter physics, the goal was to ensure that in each clock cycle only the target electron wave packet carrying a valid signal reaches the downstream energy detector, while low-energy background and high-energy stray wave packets are reflected before reaching the detector.
Hy4 preview autonomously completed the construction of a quantum scattering solver, robust optimization of a five-barrier structure, and independent time-dependent evolution verification, reducing the average leakage rate in the high-energy stopband from the baseline 48.2% to 4.8%. The model also automatically generated an interactive simulator that encapsulates the full solving workflow and supports parameter tuning and re-solving—autonomously organizing and executing the complete pipeline from multi-stage scientific computation and result verification to the delivery of an explorable artifact.
A Classic Geometry Problem
Working with Hyra, Hy4 preview achieved a major breakthrough on the century-old classic geometry problem—the three-dimensional Blaschke–Lebesgue problem—pushing the lower bound on volume from 0.380799 to 0.41104 in one leap. Compared with the 0.41986 given by the Meissner tetrahedron conjecture, this result leaves only the last 2% gap before the conjecture's final proof.
* Full proof available at: https://github.com/Tencent-Hunyuan/Hyra-results/blob/main/AI4Science/3d_blaschke_lebesgue/3d_Blaschke_Lebesgue.pdf
Hy4 preview is now open-sourced, available on Tencent Cloud TokenHub and OpenRouter, and can be experienced on products such as WorkBuddy/CodeBuddy, Yuanbao, and ima. We remain committed to openness and cost-effectiveness, so that model progress benefits more users.
Hy4 preview is an early version in the Hy4 iteration cycle. There is still considerable room for improvement in both pre-training and post-training, and there are known issues, such as prolonged deliberation on complex tasks and a tendency toward excessive self-verification. We will continue to iterate rapidly. As with Hy3 preview, we hope that releasing Hy4 preview as soon as possible will bring in broad real-world feedback to significantly improve the official Hy4 release. At the same time, we will continue to leverage our unique advantage of deep collaboration with Tencent products and experts to make productivity gains ever more accessible and to keep raising the ceiling.
Open-source links:
• HuggingFace: https://huggingface.co/tencent/Hy4-preview
• Github: https://github.com/Tencent-Hunyuan/Hy4-preview
• Modelscope: https://modelscope.cn/models/Tencent-Hunyuan/Hy4-preview
• Gitcode: https://ai.gitcode.com/tencent_hunyuan/Hy4-preview
Try the model:
• https://aistudio.tencent.com/
Hy Blog:
• https://hy.tencent.com/research/hy4-preview
Or click "Read More" to try the model directly.