Tencent Hunyuan Releases Hy4 Official Version, Introducing Vision and Computer Operation Capabilities for the First Time

@@@space height = 16 @@@

@@@buttonGroupCard

[[buttons]] text = "Hy AI Studio" icon = "hunyuan" url = "https://aistudio.tencent.com/"

@@@

@@@space height =32 @@@ Today, we are releasing the official version of Hy4. With another comprehensive leap in model capabilities, Hy has joined the ranks of the strongest open-source models for the first time.

Compared with Hy4 preview, Hy4 has significantly expanded the scale of pre-training and post-training, introduced visual capabilities for the first time, and greatly improved comprehensive capabilities such as instruction following, long-document understanding, long-horizon tasks, and deep research.

# Continuously Unlocking Real Productivity The comprehensive improvement of foundational capabilities, along with continued collaboration with numerous products and businesses within Tencent, enables Hy4 to unlock more real-world productivity scenarios.

## Vision and Computer Tasks Hy4 introduces visual capabilities for the first time, and can combine perception, reasoning, and tool calling to perform tasks such as fine-grained image processing and complex chart analysis. Hy4 also supports computer operation for the first time, and can flexibly combine the use of the command line and graphical interfaces. We conducted specialized training for professional software such as Blender and CAD and for long-horizon cross-application tasks, continuously optimizing the planning and execution of professional tasks.

@@@

@@@space height = 56 @@@

## Front-end and Game Development The introduction of visual understanding has also significantly improved Hy4's front-end and game development capabilities. We carried out targeted optimization for complex front-end interactions, business logic, state management, front-end/back-end integration, and other aspects. The interfaces generated by the model are not only good-looking but also fully functional and usable. In game development scenarios, we collaborated with professional game developers at Tencent IEG to continuously optimize data and evaluation quality, with a focus on improving capabilities such as feature implementation, gameplay logic, state understanding, and asset processing. We achieved deep adaptation to mainstream engines such as UE, Unity, and Godot, and began expanding to multiplayer online game development. Compared with Hy4 preview, Hy4 improved by 19.1% and 9.8% on the GameDev and GameLogic evaluations, respectively.

[[items]] title="Electric Flying Car" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/c9e0c958/2026110402272082_97d8e3f11949773781b3140fd95ce9ce.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793730452;2109090452&q-key-time=1793730452;2109090452&q-header-list=&q-url-param-list=&q-signature=e04d7a0867e6a9a420d5d461757274d5ecacaec5" description="Prompt: Generate a web page with the Tesla Roadster launch event's 'vertical loop' test as the visual core. The web page is not a traditional vertical landing page, but an interactive test bench that fills the browser viewport, featuring a nighttime test site 3D scene, realistic vehicle dynamics, multiple selectable vehicle models, broadcast telemetry HUD, and crash wreckage simulation.\nPlease focus on recreating:\n Nighttime launch event test site: floodlight towers + volumetric light beams, LED big screen, grandstand phone lights, distant mountains and horizon lights\n Straight track → clothoid → R 10.5 m circular arc, ending in a 170° steel truss vertical loop\n Five selectable vehicle models, default Tesla Roadster (2026 production design)\n Procedurally generated high-fidelity car models calibrated against official side/top views\n W to move forward / S to reverse, realistic longitudinal dynamics and grip limits\n Real-time calculation of normal force at every step: if speed is insufficient, the car detaches and falls; if it enters the curve too fast, the structure is overloaded\n Roadster: fan suction holds it upside down at 170°; after cresting, cold-gas thrusters flip it upright + decelerate and land\n XPENG AEROHT Land Aircraft Carrier: after detachment, the rear bay opens, the flight body rises and deploys six rotors for a stable landing\n Insufficient power or handling errors cause a crash: wreckage tumbles, wheels detach, debris flies, fire and continuous black smoke\n Broadcast telemetry-style HUD: speedometer, downforce, pitch, altitude, g-force, attitude indicator, attachment window curve\n Curve entry speed limit prompt and overspeed red warning\n Overall premium and restrained, like a real launch event livestream, not a cartoon game UI\nI. Page Structure\nThe entire website has only one full-screen 3D scene, and all UI is a HUD layer floating above the scene:\n Loading overlay (BUILDING TEST SITE)\n Main scene: nighttime test site + vertical loop + vehicle\n Top bar, left column (vehicle selection / parameters / tasks), right column (telemetry), bottom instruments\n Result card (PASS / CRASH) and 'Principles' popup layer\nDo not create a traditional Header + Hero + Section + Footer web page.\nKeep the entire experience as:\n100vw × 100vh\ncanvas: position: fixed; inset: 0\nhtml, body: overflow: hidden\nKeyboard and touch are used only for driving and switching, and do not produce any page scrolling." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110320012269_711abfd18182d627066f8ecaf5069256.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793707284;2109067284&q-key-time=1793707284;2109067284&q-header-list=&q-url-param-list=&q-signature=d57469d03ad81577dd974b3b79940e3f32aee0fd"

[[items]] title="QQ Arcade" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/c9e0c958/2026110402305312_aba05a75e2c776106a9e75f30324fed5.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793730671;2109090671&q-key-time=1793730671;2109090671&q-header-list=&q-url-param-list=&q-signature=87a5fa394693957a7c09d943c4c1e6f3d20a648d" description="Prompt: Please create a genuinely playable 'QQ Happy Arcade Hall': put the 2000s QQ game lobby into an exquisite 3D arcade room. When opened, first show the complete lobby; click a round stool to take a seat, the camera pushes in toward the cabinet, and then you select and play mini-games on the screen. It should have a nostalgic atmosphere, rich visuals, and complete interaction—not a static showcase page.\n\n1. Overall Visuals\nThe theme is 'QQ game memories from 2003': lively, friendly, bright, with a touch of toy-model and miniature-scene texture.\nThe room uses cream-yellow walls, a deep-blue ceiling, blue-and-yellow checkerboard floor tiles, warm yellow lighting paired with blue, orange, green, and purple cabinets.\nThe cabinet bodies have rounded chamfers, thick shells, glossy paint highlights, and a small amount of chrome trim. The floor has soft reflections, and objects cast contact shadows where they meet the ground.\nLight boxes and neon glow moderately, but text, screens, and cabinet colors remain clear. Do not make it a dark cyberpunk room or a modern minimalist interface.\n\n2. Lobby and Cabinets\nThe default camera faces the room with a slight downward angle, showing the floor and the depth of the cabinets. The central main cabinet is the largest and most eye-catching, with two side cabinets on each side, and all five screens visible.\nThe central main cabinet uses a blue-and-white color scheme: a 'QQ Games' light box on top, a thick screen bezel with a 4:3 screen, a protruding control panel, a red joystick, colorful physical buttons, a coin slot below, and an orange foot base. The shape should have layers; do not just assemble a cabinet from rectangular boxes.\nOn top of the main cabinet sits a penguin statue: black-and-white body, creamy white belly, yellow beak and feet, red scarf, eyes closed in a smile, one flipper raised waving and the other hanging naturally. The flippers are flat, thin sheets with soft contours.\nThe four side cabinets are respectively orange QQ Tang, green Lianliankan, a purple music rhythm machine, and green Dou Dizhu. Each has its own light box, dynamic standby screen, coin slot area, and round stool.\nThe stools use blue-purple soft cushions and metal legs, and give clear seat-available feedback when the mouse hovers over them.\nThe room includes a 'QQ Happy Arcade Hall · 2003' sign, pendant lights, balloons, a penguin figurine shelf, a claw machine, a friends-online leaderboard, a today's experts leaderboard, and game posters. The furnishings are rich but have a clear hierarchy and do not block the main cabinet screen.\n\n3. The Game Lobby on the Screen\nUse the visual language of early desktop clients: blue gradient title bar, menus and toolbars, small Songti text, thin borders, glossy buttons, and orange event banners.\nThe layout includes game categories and user profile on the left, game zones and dynamic previews in the middle, a table seat list below, and friends online and chat messages.\nThe content is dense but orderly; do not fill the screen with four large cards.\nThe previews of the four games show real gameplay animations. Clicking a game entry, the preview area, or an empty seat can start; the keyboard can also select and confirm.\nPause, help, and settlement windows keep a consistent style. The settlement includes a penguin illustration, result title, score, and replay button.\nThe screen displays real-time game content; after taking a seat, text is readable and clicks are accurate.\n\n4. Four Playable Mini-Games\nQQ Tang:\nA warm-toned cartoon grid map, with tiles and obstacles that have thickness. Arrow keys move, space places a bubble, and the bubble explodes in four directions after a delay, can destroy obstacles, and can trap characters. Computer opponents move, place bubbles, and dodge; win/loss settlement and restart are supported.\n\nLianliankan:\nThe board uses penguin, emoji, and small prop patterns. Click two identical patterns; the connecting line can turn at most twice and can route around the outside of the board. Includes a countdown, combos, hints, reshuffle, and settlement; the opening board is solvable, and when no moves remain it can be restored.\n\nMusic Rhythm Machine:\nThe screen includes a stage, dancers, a direction dance-step sequence box, and a beat bar. Input directions in order, press space to confirm on the beat, and display PERFECT, GOOD, MISS, and combos. Music and judgment are synchronized; do not make it a four-lane falling-note interface.\n\nDou Dizhu:\nThe player faces two computer opponents, including dealing, bidding for landlord, selecting cards, playing cards, passing, hints, and settlement. Card types and rank relationships are correct, and the opponents can play a full round. The player's hand is spread horizontally, and the opponents' card backs are stacked vertically.\n\nAll four games can be entered from the lobby, played to settlement, and then return to the lobby. You may control the scope of content, but it cannot be only animations or button feedback. V. Camera, Controls, and Sound Click the main unit's round stool, and the camera smoothly pushes in to a position suitable for reading the screen; click the side unit's round stool to enter the corresponding machine and start the corresponding game. After exiting, it smoothly returns to the lobby. Keyboard operations are linked with the physical controls: the joystick tilts with the direction, and buttons sink and light up when pressed. Clicking the controls with the mouse also triggers the corresponding actions. Provides operation guide, return to lobby, fullscreen, mute, and model viewing entry points. Model viewing lets you observe around the main unit. The lobby and the four games each have different looping music, along with sound effects for clicking, sitting down, clearing, placing bubbles, beat judgment, and playing cards. Sound only plays after the first operation; when switching music, it fades in and out without overlapping playback. Esc pauses, switching away from the page pauses automatically, and on resuming it does not jump around or fire repeated inputs.

VI. Implementation and Delivery Use a true 3D scene. Three.js is recommended for interaction and display, Blender for making models and baking lighting, and Canvas for drawing the real-time interface and games inside the machines. The final delivery is a single-file HTML with the data needed for models, textures, and audio embedded, runnable by double-clicking in a desktop browser, with no dependence on login, databases, or external resource requests. Use 1280×720 as the primary viewing size, and adapt to narrower windows. After scaling, text does not overlap and operation entry points remain usable. First get "lobby → sit down → select game → play → pause → settlement → return" working end to end, then refine the four games, and finally polish models, lighting, furnishings, effects, and audio. Before delivery, actually open the final HTML and check that the five screens display properly, all four games can be played to completion, clicks are accurate, the camera does not clip through models, and the console has no errors. Provide a runnable finished product, along with screenshots of the lobby panorama, main unit seating, and the four games. Directly state any parts that were not verified.

Start making it now. First give a brief implementation order, then directly push forward the first runnable version. Design unspecified furnishings, copy, and details yourself, keeping the overall style consistent.

@@@

@@@space height = 56 @@@

## Financial and Office Scenarios We built thousands of workspaces and tens of thousands of high-quality tasks based on real financial scenarios, so the model can fully learn the professional processes and norms in fund research, securities analysis, quantitative strategies, asset allocation, corporate finance, tax and accounting, and other directions. Together with Tencent FinTech FiT, we conducted a financial crowdsourced test with 213 experts, of which 153 came from external leading financial institutions, and 70% have more than five years of work experience. Based on the crowdsourced test of 579 real financial tasks, Hy4 ranks first among open-source models in overall performance (Hy4: 3.04/4; DeepSeek-V4.1-Flash-1017: 2.97/4; Kimi K3.1 Pro: 2.68/4; GLM-5.3: 2.68/4). In addition, in the internal evaluation IRAB of the financial agent company Xuntu Technology, Hy4 improved by 12.03 percentage points over Hy4 preview. To improve the practicality of financial and other office scenarios, we also launched targeted search optimization for time-sensitive retrieval, multi-source information integration, and factual consistency verification. In our own tests, Hy4's hallucination rate dropped from 15.76% in Hy4 preview to 5.45%, and the error rate caused by guessing dropped from 23.1% to 5.6%. In addition, Hy4 has focused on improving human-machine co-writing capabilities in office scenarios, supporting users and Hy to collaboratively edit Word, Excel, PPT, and other office files in real time. Blind evaluation shows that Hy4 has significantly improved over Hy4 preview in document aesthetics, stability, and usability. The PM challenge turns the Copilot panel red and answers the two challenge questions using the assumptions in the model.\n\nAdditional Instructions: Use only plain HTML/CSS/JS, no frameworks, and no ES modules; double-clicking index.html should run it. All data is hard-coded sample data, and all UI text is in English. Color scheme: background #04060B, overlaid with a faint 48px grid and two glows in blue and purple; cards are glassmorphic with 14px rounded corners; accent colors are #5B8CFF, #38E1FF, #9B6BFF, green #22D39A, red #FF5470, amber #FFB547; numbers use a monospace font. The business logic must be internally consistent: operations depend on one another (prep → notes → model → export → PM), and when a later step is clicked early, first ask whether to run the missing prerequisite steps; the clock can only move forward; only one workflow can run at a time, with a Skip button to fast-forward; the values of key drivers follow only the saved model version; the exported Excel is a snapshot of that moment, and is marked stale when TP deviates by more than 5% or the rating changes. Valuation is based on the embedded Excel's Valuation sheet: the target price is the weighted average of SOTP, EV/Sales, and DCF by weight, then rounded. The footer states: "Illustrative data for demonstration only. Not investment advice."\n\nThe valuation model is also built as a real Excel workbook, embedded in the Model view: it looks like real Excel (white grid, column letters and row numbers, frozen headers and first column, sheet tabs at the bottom, formula bar at the top; number formats such as thousands separators, negative numbers in parentheses, percentages, and multiples with an x; blue text for inputs, black text for formulas, green text for cross-sheet references). There are 7 sheets in total: Cover, Drivers (revenue split by business line), IS, BS, CF, Valuation (WACC, year-by-year DCF plus terminal value, reverse DCF, SOTP, weighted target price, scenarios), Estimate Changes; columns are FY2022–FY2025 historical plus FY2026E–FY2032E forecast, with thousands of formulas, and the three statements tie out year by year (CF ending cash flows back into BS, and BS has a Balance check row each year). Only input cells can be changed; after pressing Enter the whole workbook recalculates, and changed cells flash and highlight; clicking any formula cell highlights the reference chain from inputs to it in sequence, and the sidebar lists the value at each step; the header permanently shows tie-out checks (balance sheet balances, cash is consistent, and the target price matches the Model view); one click exports a .xlsx that preserves formulas, formatting, and frozen panes, and when opened in Excel the values are exactly the same. The formula engine is implemented from scratch in pure JS, with no libraries.

## Scientific Research and Model Development Hy4 combines complex reasoning, scientific programming, and long-horizon task execution to further enhance its ability to advance scientific research end to end. In the internal TenureTrail benchmark covering 7 major disciplines and 123 executable research tasks, Hy4 demonstrated solid research capabilities in areas such as mathematical construction, physical modeling, Earth and environmental science, systems engineering, and socioeconomic decision-making, ranking first among open-source models. Hy4 is also preparing for recursive self-improvement (RSI), improving capabilities across various model development tasks. Combined with the [Hyra 2.0](https://hunyuan.tencent.com/research/hyra2) scaffold, Hy4 optimized the pretraining infrastructure for internal small models of the same architecture and automatically trained the Jev model, which outperforms the official version on multiple benchmarks. At the same time, we have also open-sourced HyCell, a single-cell representation foundation model, and HyPDE, a physical field foundation model, developed by Hy4 in collaboration with Hyra2.0, hoping to turn the results of model-assisted scientific research into tools reusable by the community.\n\n## Long-term Personal Assistant Hy4's in-context learning ability continues to improve, enabling it to better understand complex information from multiple sources. With the WorkBuddy enterprise WeChat suite, Hy4 can connect to daily work information such as chat records, emails, documents, schedules, and meetings. When multiple projects are advancing in parallel, Hy4 can clarify each project's history and latest progress, identify tasks to follow up on, meetings to prepare for, and items approaching deadlines, promptly remind users, and help prioritize. In addition, the team collaboration feature powered by Hy4 will launch in Workbuddy with the new version of "Projects".

@@@testimonialSlider ariaLabel = "Customer Testimonials" itemsPerPage = 2 mobileItemsPerPage = 1 loop = false defaultActivePage = 0 showCount = true countSeparator = " / " countMinDigits = 2 countLabelTemplate = "Page {current} of {total}" emptyText = "No testimonial content available" contentEmptyText = "No testimonial content available" previousLabel = "Previous testimonial" nextLabel = "Next testimonial"

@@@

# Responsible Open Source and Release We are fully aware of the potential safety risks posed by the model's cyber and biological capabilities, and are committed to training, open-sourcing, and releasing it in a more responsible manner.

During the Hy4 training phase, we included only defensive cybersecurity data related to PoC generation and vulnerability remediation, while excluding offensive training content such as vulnerability exploitation, reverse engineering, and penetration, striving to avoid improvements in potential offensive capabilities. We organized internal security experts from Tencent Xuanwu and Keen Labs to conduct audits focusing on cryptanalysis, reverse engineering, vulnerability exploitation, and other areas. The results show that Hy4 does not exhibit prominent risks in offensive capabilities such as reverse-solving cryptographic algorithms, countering binary obfuscation, and vulnerability exploitation.

In terms of biosecurity, Hy4's open-ended experimental design capability has not yet reached the high-risk threshold (Hy-BioExperimentBench: 30.4/100). We organized internal experts to audit the model's response trajectories and found that the model has a solid grasp of basic biological knowledge, but its ability to translate knowledge into executable experimental protocols—a key step—remains limited, and it has not been found to significantly lower the threshold for carrying out biologically harmful acts.

Hy4 is now available on Tencent Cloud and OpenRouter, and can be experienced on products such as WorkBuddy/CodeBuddy, Yuanbao, and ima. Its performance significantly exceeds Hy4 preview while maintaining the same API price.

@@@space height = 32 @@@

::: center Hy4 API Pricing (per million Tokens) ::: ::: center |Input (cache hit) | Input | Output | | ------ | ------ | ------ | | 0.3 yuan | 6 yuan| 18 yuan | :::

@@@space height = 32 @@@

We plan to open-source the model weights, and have partnered with the vLLM/SGLang community and Inferact Inc. to provide open-source ecosystem support. We will also release a model system card at that time, including more comprehensive details such as safety evaluations.

# Limitations and Future Outlook Although Hy4 is already one of the strongest open-source models currently available, a considerable gap remains between it and top closed-source models, especially on the highly difficult test sets we built internally in areas such as frontier mathematics, long-horizon development, model iteration, and visual question answering. We will continue to steadily scale up pre-training and post-training, striving to narrow the gap with top closed-source models.

On the other hand, we plan to release Hy4 flash in the near future, which will significantly improve the model's cost-effectiveness and inference speed, and enable more frequent collaborative iteration with products to continuously improve practicality.