腾讯发布 Hy4 大模型并预告 Hy4 Flash
腾讯正式发布新一代大模型 Hy4,同步上线腾讯云 TokenHub 与 OpenRouter 平台,并公布多项基准测试结果。官方同时预告将推出主打高性价比与更快推理的 Hy4 Flash,强调与产品侧更紧密的协同设计。
报道正文
@@@coverCard type = "video" date = "November 4, 2026" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110219294526_20bbdf2a313c4cad81b3fdb22932e7ff.mov?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793618987;2108978987&q-key-time=1793618987;2108978987&q-header-list=&q-url-param-list=&q-signature=7b331fb3eec1fdd37e468ac2348a8fdfd77b1406" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110219302736_cb6bad8aa277cf9d40dc98cfa2d4ad3a.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793619030;2108979030&q-key-time=1793619030;2108979030&q-header-list=&q-url-param-list=&q-signature=150aa51ca6edd06bbe479a9f85ee8bab9b9ade49" image_alt = "hy4PreviewBg" title = "Introducing Hy4" aspectRatio = "" mediaWidth = "100%" mediaMarginTop = "32" mediaMarginBottom = "64" videoLoop = 0 @@@ @@@space height = 16 @@@ @@@buttonGroupCard [[buttons]] text = "Tencent Cloud" icon = "tencentcloud" url = "https://console.tencentcloud.com/tokenhub/models/detail?modelId=hy4" [[buttons]] text = "OpenRouter" icon = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026070622081339_db7b4e83c4ccd253de47d27e656c6a1f.svg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1783346895;2098706895&q-key-time=1783346895;2098706895&q-header-list=&q-url-param-list=&q-signature=44a94d95270276ae213ffe4696b88387477b9592" url = "https://openrouter.ai/tencent/hy4" [[buttons]] text = "Hy AI Studio" icon = "hunyuan" url = "https://aistudio.tencent.ai/" @@@ @@@space height =32 @@@ Today, we’re releasing Hy4, another major leap forward that places Hy among the strongest open-source models for the first time. Compared with Hy4 preview, Hy4 substantially scales up both pre-training and post-training. It adds vision capabilities for the first time and delivers major gains in instruction following, long-context understanding, long-horizon tasks, deep research, and more.
# Unlocking real-world productivity With broad improvements in core capabilities and ongoing co-design with product teams across Tencent, Hy4 moves beyond benchmarks to open up more real-world use cases. ## Vision and computer use Hy4 supports vision perception for the first time. By combining perception, reasoning, and tool use, it can understand fine-grained image details and interpret complex charts. Hy4 also supports computer use for the first time, working across command-line interfaces (CLI) and graphical user interfaces (GUI). Through targeted training on specialized software such as Blender and CAD, as well as long-horizon workflows that span multiple applications, we significantly improved its ability to plan and carry out professional tasks. @@@mediaGallerySwiper [[items]] title="Locate a rooftop component approximately 2 meters in size in a 2.8-gigapixel aerial image" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412350446_2f755945ab99ef0048c13fe3df876f23.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793766906;2109126906&q-key-time=1793766906;2109126906&q-header-list=&q-url-param-list=&q-signature=8f64d85cccde880b1b48e7d1937122a440928ec2" description="Prompt: Find the easternmost white square component in the solar panel area to the upper right of the warehouse with “KLOOSTERBOER” written on its roof." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412372735_9f360c5ab7736510df54c882e9dbf188.jpeg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767050;2109127050&q-key-time=1793767050;2109127050&q-header-list=&q-url-param-list=&q-signature=a10c7435514a4b443e9ec7363e5332ac3f751a44" [[items]] title="Using multiple tools to create a DIY home improvement program" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412392226_8c6cea5f8cc23b5d6912d4c49e7d4972.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767167;2109127167&q-key-time=1793767167;2109127167&q-header-list=&q-url-param-list=&q-signature=10e4db28e34effcd1f5b4c3a4b9814760ae12464" description="Prompt: Use this blueprint as a reference to rebuild the entire building's interior space in Blender with MCP. You can use CC0 assets from the plugin, and if you can't find what you need, make it yourself. Try to create materials, lighting, colors, and structures as realistically as possible. Export the Blender scene with models and textures, then import it into UE, keeping the scene layout and lighting. Before exporting, merge models that should be merged, like combining floor tiles into one and merging plants in pots, to simplify the import/export process and ensure engine optimization. In the UE level, add a first-person preview mode where you can select key interior elements like walls or tables and chairs and change their materials. The end result is that players can freely DIY the interior decoration of the room." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110320245856_df72b7ce4b0fc63bf4b0ca0fe1f021c8.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793708698;2109068698&q-key-time=1793708698;2109068698&q-header-list=&q-url-param-list=&q-signature=b6491f4115a63bc33bec45a4c7907de10b216060" [[items]] title="Reassembling a 900-piece Jigsaw puzzle to reveal Monet’s Bordighera" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412411758_7325d50a2ba30f0bc3554f4072af3c0a.mov?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767284;2109127284&q-key-time=1793767284;2109127284&q-header-list=&q-url-param-list=&q-signature=7d82eb908ccb2978ca7214504cfe0ce6cae55751" description="Prompt: Image path: /Users/hunyuan/demo/input_jigsaw_puzzle_30x30_monet.png\n\nUsing only this image, reassemble the 900 pieces into a complete 30 × 30 jigsaw puzzle. Do not search the web or look for the original complete image or a solution in local files. Treat the directory containing the input image as read-only and save all generated files in your workspace. Briefly outline your plan, then proceed directly with execution. You may write scripts.\n\n1. Segment and assign IDs to the pieces at the input image’s original resolution. Verify the piece count and preserve each piece’s original pixels, contour, and location in the input image. Maintain the original proportions, apply the same geometric transformations to each piece’s image and mask, and handle RGB/BGR conversions correctly to avoid color shifts.\n\n2. Validate a representative sample of processed pieces and initial joins before expanding the assembly. Consider both visual continuity between neighboring pieces and the overall layout, retaining alternatives for ambiguous matches. When conflicts arise, revisit earlier joins and backtrack, splitting assembled groups if necessary. Reuse validated results; if progress stalls, identify the cause rather than blindly repeating computations.\n\n3. Before blending or retouching seams, independently verify that every piece is used exactly once, the grid is complete, and there are no conflicting placements. Check all adjacencies programmatically, inspect the full image and suspicious seams, and correct any issues. A complete grid alone does not establish that the assembly is correct.\n\nDeliver a full-resolution PNG, a preview image, and each piece’s final grid position (row and column). Display the reconstructed image and confirm that its colors and overall orientation are correct. Clearly identify any placements that remain uncertain." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412412436_5fb6b9fb09f1795dc718ea6dcb234c98.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767285;2109127285&q-key-time=1793767285;2109127285&q-header-list=&q-url-param-list=&q-signature=e10f11c6b10b5795272ace8c26087c03a139fe3b" [[items]] title="Visual geolocation and pinpointing" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412445566_0454178fc24779d62779d54189cc09b6.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767508;2109127508&q-key-time=1793767508;2109127508&q-header-list=&q-url-param-list=&q-signature=515c06dca8f2c8cde598c24ef5378b3b440ded6b" description="Prompt: What city is this? Where was this photo taken?" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412450872_1b65d98f3c4a2af27230ed9dbb91419e.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767509;2109127509&q-key-time=1793767509;2109127509&q-header-list=&q-url-param-list=&q-signature=c0e453bd19f8b165b5426ed336ee1a8212136993" [[items]] title="Hy stickers — drawn with a mouse in JS Paint" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412465138_06a7dcecc91817b5110b719a51ad9721.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767614;2109127614&q-key-time=1793767614;2109127614&q-header-list=&q-url-param-list=&q-signature=708a7333da7cf1849e5b90bda0c72aa8438a4a4d" description="Prompt: Use JS Paint (jspaint.app) in Chrome to draw hy.png only with the mouse, and record each mouse action and screen change as a video. Requirements: Playwright records the entire screen and displays the cursor; drag the canvas to 512x460; Lab clustering uses 28 colors including white, and adjusts colors only by clicking on the color gamut and brightness bar in Edit Colors; use a square brush with 8/5/2px three times from thick to thin, read back the canvas each time, and merge into straight brush strokes by gaining on the grid; JS only reads and does not write to the canvas. Output output/: entire process at 8x speed" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412465415_7009194ce5c2541529e10465a0fb40e6.jpeg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767615;2109127615&q-key-time=1793767615;2109127615&q-header-list=&q-url-param-list=&q-signature=698fb507c74ec698278d87a70e6f86149c12caa0" @@@ @@@space height = 56 @@@ ## Frontend and game development Integrating visual understanding has driven a step change in Hy4’s frontend engineering and game development capabilities. For frontend work, we optimized Hy4 for complex UI interactions, business logic, state management, and full-stack integration, so that generated interfaces are not only visually polished but also functionally complete and production-ready. For game development, we partnered with professional developers at Tencent Interactive Entertainment Group (IEG) to curate high-quality training data and build evaluations. We focused on feature implementation, gameplay logic, state tracking, and asset pipelines. We also extensively adapted Hy4 for major engines such as Unreal Engine, Unity, and Godot, and expanded its capabilities to multiplayer game development. Compared with Hy4 preview, Hy4 improves by 19.1% on GameDev and 9.8% on GameLogic. @@@mediaGallerySwiper [[items]] title="Tencent Penguin Island Digital Sand Table" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412511919_f17bce0d9b1eeca86a7263a03ac20d93.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767885;2109127885&q-key-time=1793767885;2109127885&q-header-list=&q-url-param-list=&q-signature=b87de1e64f400a56c009547995f9b7c7cb1eb6b0" description="Prompt: Objective\n\nBuild a 3D digital scale model of Tencent Penguin Island in Three.js. Deliver a single, self-contained HTML file that opens offline by double-clicking. Embed Three.js 0.184.0 directly in the file. It must have no remote dependencies, including CDNs, fonts, or textures.\n\nModel every building, road, green space, stretch of coastline, tree, person, and vehicle within the same 3D scene. Users must be able to rotate the view freely and walk through the island at ground level. A static bird’s-eye view, a scrolling landing page, or a scene rendered as a background image will not meet this requirement.\n\nSite and architecture\n\nThe site occupies a peninsula that runs north–south through a bay. Use blue-gray water, white and silver buildings, a central green spine, and a road network that circles the island. Give particular attention to five architectural groups.\n\n1. Tengyun Center (腾云中心) — Three low-rise, cloud-like buildings with raised cylindrical volumes and steel-truss skybridges.\n2. Yunhai Towers (云海大厦) — Six staggered towers of varying heights, with parametrically modeled metal fins.\n3. Penguin Apartments (企鹅公寓) — Eleven residential blocks offset in a sawtooth arrangement, with shared stepped terraces.\n4. Tencent Helix (腾讯螺旋) — highest priority. Shape the tower from four facade surfaces. Each surface sweeps around continuously and broadens as it rises. At ground level, the facades turn inward to form J-shaped feet. The spiral must come from the geometry of the facades themselves. Do not add external rings or model the tower as four twisted ropes.\n5. Science and Technology Museum (科技馆) — A continuous ring-shaped green roof encloses an inner courtyard. Display the Chinese annotation “图像拟合、非测绘” (“Modeled from imagery, not survey data”).\n\nInteraction\n\nProvide orbit controls and ground-level exploration in both first and third person. Show a penguin avatar in third-person mode. Include weather and day/night controls, a minimap, building details, floor-by-floor cutaway views, a cinematic guided tour, screenshot capture, and video recording.\n\nInterface\n\nUse light-colored frosted-glass control panels, clear Chinese typography, and a collapsible tool overlay. Night mode should change only the 3D environment. Keep the panels light and legible after dark, and avoid a black, neon-lit cockpit aesthetic." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110320005933_0129832274e3aabfb6200975a2687bc7.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793707262;2109067262&q-key-time=1793707262;2109067262&q-header-list=&q-url-param-list=&q-signature=f945b083bac3557c5ac3689a09380ad714688c39" [[items]] title="Electric Flying Car Climb Test" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412525050_d5e4b6b53df1e281b054a61dffbbcb81.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793767973;2109127973&q-key-time=1793767973;2109127973&q-header-list=&q-url-param-list=&q-signature=6ce9681dff1bd5951a4fafc86e370aa537dc440e" description="Prompt: Build a full-screen interactive web experience centered on the vertical curve test at the Tesla Roadster launch event. The entire browser viewport should act as a driving simulator, with a 3D nighttime proving ground, realistic vehicle dynamics, five selectable vehicles, a broadcast-style telemetry HUD, and a simulation of crash damage and debris.\n\n### Scene and simulation\n\n- Recreate the proving ground at night. Floodlight towers cast visible beams through the air. A large LED screen lights the venue, phone lights dot the grandstands, and distant mountains rise above scattered lights on the horizon.\n- Build a steel-truss vertical track. A straight approach leads into a clothoid transition, followed by a circular arc with a 10.5 m radius. The track ends at 170°.\n- Offer five selectable vehicles, with the Tesla Roadster (2026 production design) selected by default.\n- Create high-fidelity procedural vehicle models. Use official side and top views to match each vehicle’s proportions and silhouette.\n- Use W to drive forward and S to reverse. Simulate realistic longitudinal dynamics and enforce tire grip limits.\n- Calculate the normal force at every simulation step. If the vehicle loses contact because its speed is too low, it leaves the track and falls. Excessive entry speed causes structural overload.\n- Tesla Roadster — Fan-generated suction holds the car against the track at 170°, nearly upside down. After it clears the top, cold-gas thrusters rotate it upright and slow its descent for landing.\n- XPENG AEROHT Land Aircraft Carrier — After the vehicle leaves the track, its rear bay opens. The flight module rises out, unfolds its six rotors, and makes a controlled landing.\n- Insufficient power or a driving error should result in a crash. Simulate the wreck tumbling, wheels breaking free, debris scattering, a burst of flame, and black smoke that continues to rise.\n\n### Telemetry and visual direction\n\n- Use a motorsport broadcast-style HUD with a speedometer, downforce, pitch angle, altitude, g-load, an attitude indicator, and a curve showing the operating range in which the vehicle can maintain contact with the track.\n- Display the speed limit for entering the curve. Show a red warning when the driver exceeds it.\n- Keep the presentation refined and restrained, with the visual language of a live vehicle launch broadcast. Avoid cartoon-style game graphics.\n\n### Page structure\n\nUse a single full-screen 3D scene with all interface elements floating above it as HUD overlays.\n\n- Loading overlay displaying BUILDING TEST SITE\n- Main scene containing the nighttime proving ground, vertical track, and vehicle\n- Top bar\n- Left panel for vehicle selection, parameters, and test objectives\n- Right panel for live telemetry\n- Bottom instrument cluster\n- Result card displaying PASS or CRASH\n- “How It Works” modal explaining the physics behind the test\n\nDo not use a conventional scrolling landing-page structure with a header, hero, content sections, and footer.\n\n### Viewport and input\n\n- Keep the experience at 100vw × 100vh.\n- Pin the canvas to the viewport with canvas { position: fixed; inset: 0; }.\n- Disable page scrolling with html, body { overflow: hidden; }.\n- Reserve keyboard and touch input for driving and switching selections. Neither should cause the page to scroll." 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="Interactive 3D Web Explainer" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412544839_0501cea2dc2851e3202424c887777e01.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768092;2109128092&q-key-time=1793768092;2109128092&q-header-list=&q-url-param-list=&q-signature=89749b6df2b48bb1d4e4c0d8165489f21d9db1db" description="Prompt: Create a polished, interactive 3D educational webpage based on this reference image. Deliver only a single .html file.\n\n# Background Requirements\nUse a starry-sky background with an educational information panel beside the model. Clicking labels on the model should display the structure’s name, function, and common exam questions. Support cutaway views, exploded views, side-by-side comparison of multiple objects, and comparison with the original reference image.\n\n# 3D Modeling Requirements\n\n## Escherichia coli (E. coli)\n- Create a rod-shaped bacterium with a cutaway opening that reveals its internal structures.\n- Model the following layers from outside to inside, using clearly distinguishable colors:\n Cell wall (opaque orange-yellow) → cell membrane (thin green layer) → cytoplasm (semi-transparent light blue).\n- Nucleoid: Create a blue DNA mass using multiple randomly intertwined TubeGeometry strands.\n- Ribosomes: Randomly distribute 100–200 small purple spheres throughout the cytoplasm.\n- Pili: Add 50–80 short tubes distributed along the surface normals.\n- Flagella: Add 3–5 long, wavy tubes extending from one end.\n\n## Cyanobacterium\n- Create an ellipsoidal cell with a cutaway view.\n- Use a thick, dark-teal outer cell wall with a granular texture, with a cell membrane underneath.\n- Use green cytoplasm containing 4–6 concentrically arranged photosynthetic lamellae. Distribute small black dots across the lamellae to represent ribosomes.\n- Build the central nucleoid region using a network of filament-like strands and several yellow granules." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412544573_d2b5ca33bd970f64a6301fa75ae2eb22.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768088;2109128088&q-key-time=1793768088;2109128088&q-header-list=&q-url-param-list=&q-signature=a46aa7e7bf7fcf5d1921c9f796566e75fc9f8f9f" [[items]] title="Creating a Left 4 Dead–Style FPS Survival Game in Unreal Engine Using Existing Assets" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412565892_7d25c1ce0ee57b4b8dc220a8f7a36840.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768224;2109128224&q-key-time=1793768224;2109128224&q-header-list=&q-url-param-list=&q-signature=5aa99683bba6fe9abf457d3389c3ecc6673abb56" description="Prompt: Create a third-person shooter game using the Hyper engine. You first implement the 3C mechanics of a TPS game in a simple scene with only a floor and four high walls. You can switch between a dagger, pistol, rifle, and shotgun using the number keys 1, 2, 3, and 4." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412570542_57da346b8c057f411d5e11b66ffaf60b.jpeg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768225;2109128225&q-key-time=1793768225;2109128225&q-header-list=&q-url-param-list=&q-signature=e5cf9d29fa53a23bb6c970560c9fcfab76c2b2f8" [[items]] title="Retro-style QQ Game Arcade" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412590065_e130560c25ccf11890b04ded835c434e.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768345;2109128345&q-key-time=1793768345;2109128345&q-header-list=&q-url-param-list=&q-signature=49441dfe217bc5382e3f7775d1140d3c5312f30b" description="Prompt: Create a fully playable "QQ Game Arcade"—a 3D, meticulously crafted arcade space housing the classic "QQ Game Hall" from the 2000s. Upon entering, the user sees the entire arcade; clicking on a round stool triggers a camera zoom toward a specific machine, where the user can then select and play a mini-game. The experience should be immersive, rich in detail, and fully interactive—far more than a static display.\n\nI. Overall Visual Style\nThe theme is "Retro Arcade: Memories of QQ Games"—lively, welcoming, and bright, with an aesthetic reminiscent of toy models or miniature dioramas.\nThe room features cream-yellow walls, a deep-blue ceiling, and a blue-and-yellow checkered floor. Warm yellow ambient lighting complements the arcade cabinets, which come in shades of blue, orange, green, and purple.\nThe cabinets themselves feature rounded edges, robust casings, a high-gloss lacquer finish, and subtle chrome accents. The floor exhibits soft reflections, with contact shadows grounding each object.\nLight boxes and neon elements glow gently, ensuring that text, on-screen content, and cabinet colors remain crisp and legible. Avoid a dark, "cyberpunk" atmosphere or a modern, minimalist interface.\n\nII. The Arcade Hall and Cabinets\nThe default camera view is positioned slightly above, looking down into the room to capture the floor space and the depth of the cabinets. A large, prominent "main" cabinet sits in the center, flanked by two smaller cabinets on either side—all five screens are clearly visible.\nThe central cabinet features a blue-and-white color scheme: a "QQ Games" light box at the top, a thick bezel surrounding a 4 aspect-ratio screen, a protruding control panel, a red joystick, colorful physical buttons, a coin slot, and an orange base. The design is multi-layered and complex, avoiding simple, boxy, rectangular shapes.\nPerched atop the central cabinet is a penguin figurine: it has a black-and-white body, a creamy-white belly, a yellow beak and feet, and a red scarf. Its eyes are closed in a smile; one flipper is raised in a wave, while the other hangs naturally at its side. The flippers are thin, flat, and softly contoured.\nThe four surrounding cabinets are themed after specific titles: an orange QQ Tang (QQ Arcade/Bomberman-style) machine, a green Lianliankan (Connect-the-Tiles) machine, a purple music/rhythm game machine, and a green Dou Dizhu (Fight the Landlord) machine. Each unit features an individual light box, an animated idle screen, a coin-insertion slot, and a round stool.\nThe stools are designed with blue-violet cushioned seats and metal legs; hovering the mouse over them provides a clear visual cue that they are available for sitting.\nThe room is decorated with a "QQ Happy Arcade · 2003" sign, hanging lights, balloons, a display shelf for penguin figurines, claw machines, a list of online friends, a "Top Players of the Day" leaderboard, and various game posters. The furnishings are plentiful yet arranged with a clear hierarchy, ensuring they do not obstruct the game screens.\n\nIII. The In-Game Lobby\nThe interface adopts the visual language of early desktop clients: blue-gradient title bars, menus and toolbars, small-sized Song-style (serif) fonts, thin borders, glossy buttons, and orange promotional banners.\nThe layout includes game categories and user profiles on the left, a central area for featured games and dynamic previews, a list of tables and seats at the bottom, and sections for online friends and chat messages.\nThe content is dense yet organized; the screen is not cluttered by oversized cards.\nPreviews for the four games showcase actual gameplay animations. Users can start a game by clicking on a game entry, the preview area, or an empty seat; keyboard navigation is also supported for selection and confirmation.\nPause, Help, and Results windows maintain a consistent visual style. The Results screen features a penguin illustration, a clear outcome heading, the player's score, and a "Play Again" button.\nThe screen displays real-time game content; once seated, text remains legible, and interactive elements are easy to click accurately.\n\nIV. Four Playable Mini-Games\nQQ Tang (QQ Bomb):\nSet on a warm-toned, cartoon-style grid map, the tiles and obstacles have a distinct, three-dimensional thickness. Players use arrow keys to move and the spacebar to place "sugar bombs." After a short delay, these bombs explode in four directions, destroying obstacles and trapping characters. AI opponents can move, place bombs, and dodge; the game supports win/loss outcomes and restarts.\nLianliankan (Tile-Matching Game):\nThe board features icons such as penguins, emoticons, and small props. Players match identical icons by connecting them with a line that can make up to two turns (paths can even go around the outside of the board). Features include a countdown timer, combo streaks, hints, a board-reshuffle option, and a final results screen. Every game starts with a solvable configuration, and the board can be reset if no moves remain.\nMusic Rhythm Game:\nThe interface includes a stage, a dancer, a sequence box for directional dance steps, and a rhythm/beat tracker. Input directional commands in sequence and press the spacebar on the beat to confirm; the game displays "PERFECT," "GOOD," "MISS," and combo counts. The gameplay is synchronized with the music, utilizing a format other than the standard four-lane falling-note interface.\n\n"Dou Dizhu" (Fight the Landlord):\nThe player competes against two computer-controlled opponents. The game includes dealing, bidding for the role of "landlord," selecting cards, playing cards, passing, receiving hints, and final scoring. Card combinations and their relative values are implemented correctly, and opponents are capable of playing through an entire round. The player’s cards are displayed horizontally, while the opponents' cards are shown as vertical stacks of face-down cards.\n\nAll four games are accessible from the main lobby; players can complete a full session—from start to final scoring—and then return to the lobby. While the scale of the content can be managed, the experience must go beyond simple animations or button-press feedback.\n\nV. Camera, Controls, and Audio\nClicking the round stool in front of the main console smoothly moves the camera to an optimal viewing position; clicking a stool by one of the side consoles launches the corresponding game. Exiting a game returns the camera smoothly to the lobby view.\nKeyboard inputs are synchronized with the on-screen physical controls: joysticks tilt in the corresponding direction, and buttons depress and light up when pressed. Clicking these controls with the mouse triggers the same actions.\nThe interface includes options for an operation guide, returning to the lobby, toggling full-screen mode, muting audio, and a "model viewer" mode that allows for 360-degree inspection of the console.\nThe lobby and each of the four games feature unique, looping background music, complemented by sound effects for clicks, sitting down, clearing elements, popping bubbles, beat-based timing, and playing cards.\nAudio begins only after the first user interaction. Music tracks transition via fade-in/fade-out, with no overlapping playback. Pressing the "Esc" key pauses the game; switching away from the browser tab also triggers an automatic pause. Upon resuming, the game state remains stable, with no erratic jumps or unintended rapid-fire inputs.\n\nVI. Implementation and Delivery\nThe project utilizes a fully 3D environment. It is recommended to use Three.js for interactivity and rendering, Blender for modeling and light-baking, and the HTML5 Canvas API for rendering the real-time interfaces and gameplay elements within the consoles.\nThe final deliverable shall be a single HTML file with all necessary models, textures, and audio data embedded. It must be executable in a desktop web browser via a simple double-click, requiring no user login, database, or external resource requests.\nThe primary design resolution is 1280×720, with responsive adaptation for narrower window sizes. Text must not overlap when scaled, and all interactive controls must remain functional. First, establish the core flow: "Lobby → Seating → Game Selection → Gameplay → Pause → Results → Return." Next, refine the four individual games. Finally, polish the models, lighting, decor, visual effects, and audio.\nBefore delivery, open the final HTML file to verify that all five screens display correctly, each of the four games is fully playable, click registration is accurate, the camera does not clip through geometry, and the console shows no errors.\nProvide a functional final build, along with screenshots of the lobby (panoramic view), the seating area, and each of the four games. Clearly note any components that have not been verified.\n\nBeginning production now. I will outline the implementation sequence briefly, then proceed directly to developing the first functional version. For any unspecified decor, text, or details, I will create designs that ensure a consistent overall style." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110412590519_442f8fd6ef494e2d1a30d10b4ae2e0e3.jpeg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768346;2109128346&q-key-time=1793768346;2109128346&q-header-list=&q-url-param-list=&q-signature=c02aed23eccae7991c0ddd4f7d85ffc333480b7f" [[items]] title="3D online multiplayer web game" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413005715_af35c4376bb63faa2aa39de2a31e182e.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768467;2109128467&q-key-time=1793768467;2109128467&q-header-list=&q-url-param-list=&q-signature=e733a5d4c929f81a94e523ec7ae7fa2e993ed6d1" description="Prompt: Making a two-player cooperative online shooting game with Three.js: two capybaras each ride a small boat, and the core gameplay is shooting monsters with guns while using the recoil from the handheld guns to push the boats. The scene is a stylized shallow sea in a 100m sailing area, set on a sunset-lit island with coral, seaweed, schools of fish, and thorny plants rooted in the seabed. The whole experience is about continuously surviving and fighting monsters, with merchant ships appearing periodically where you can buy weapon upgrades, and finally taking down the boss. Each player has 5 hearts, and if you go down, your teammate has to revive you. The server runs at 60Hz simulation, with snapshots synchronized with client prediction. The camera is top-down, and the final output is a single HTML file." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/c9e0c958/2026110402331028_5c415d3e085bf59e329866efc4e0911d.jpeg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793730793;2109090793&q-key-time=1793730793;2109090793&q-header-list=&q-url-param-list=&q-signature=9feb25f121a3b9f154602b29016fb995d9f4efdc" @@@ @@@space height = 56 @@@ ## Finance and office workflows To ground Hy4 in real-world financial work, we built thousands of realistic workspaces and tens of thousands of tasks spanning fund research, equity analysis, quantitative strategies, asset allocation, corporate finance, and tax accounting. Working with Tencent Financial Technology (FiT), we evaluated Hy4 with 213 finance professionals, including 153 experts from leading financial institutions. More than 70% of these experts have over five years of industry experience. Across 579 real-world financial tasks, Hy4 ranks first among open-weight models (Hy4: 3.04; DeepSeek-V4.1-Flash-1017: 2.97; Kimi K3.1 Pro: 2.68; GLM-5.3: 2.68). On IRAB, an internal benchmark from financial AI agent company RabyteTech, Hy4 also scores 12.03 percentage points higher than Hy4 preview. To make Hy4 more reliable for financial and office workflows, we made targeted upgrades to search, focusing on time-sensitive retrieval, multi-source synthesis, and factual consistency verification. In internal evaluations, Hy4 cut the hallucination rate from 15.76% (on Hy4 preview) to 5.45%, and reduced speculative errors from 23.1% to 5.6%. We also improved Hy4’s ability to co-edit Word, Excel, and PowerPoint files with people in real time. In blind evaluations, Hy4 shows clear gains over Hy4 preview in document formatting, reliability, and overall usability. @@@mediaGallerySwiper [[items]] title="Financial Valuation Modeling & Research Report Generation" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413051855_70a8cb4c5301bf52c28a31699bc7755d.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768720;2109128720&q-key-time=1793768720;2109128720&q-header-list=&q-url-param-list=&q-signature=3ffd19ecb3f1d8992c64b75fb865fcb477beea67" description="Prompt: use global-equity-research and neodata to build a financial model and research report on Tesla. Deliver both the Excel model and the matching PDF report. Pay close attention to your valuation logic and to where every number comes from. You can search for and draw on research from multiple institutions. List the research firm as HY RESEARCH." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413054286_13f7730993133e2e8bc6a2152e6e1707.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768742;2109128742&q-key-time=1793768742;2109128742&q-header-list=&q-url-param-list=&q-signature=e8985ce043feb998048311fa63ca0ba7790a71e3" [[items]] title="Buy-side Research Workstation for Secondary Markets, built around how analysts actually work" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413073480_7c8638bf7a7cd916eceb5002ae8d1670.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768857;2109128857&q-key-time=1793768857;2109128857&q-header-list=&q-url-param-list=&q-signature=41ab99608a38064686982b0d2adced21d21851ba" description="Prompt: Build an interactive HTML app "Analyst Workstation", with HY RESEARCH WORKSTATION written in the top-left corner: a dark, terminal-style 1920x1080 dashboard (scaled proportionally to the window) that walks a buy-side analyst covering Tesla (TSLA) through one trading day, from the 07 pre-market check to the 18 investment committee (IC). Layout: a top bar (logo, Ctrl+K command palette, Market/Call/Model/Outputs/Settings view switcher, US Eastern clock, "IC meeting in hh:mm" countdown); a left rail (last price, % change and source tag for 10 EV stocks, with a workspace file tree with NEW/UPDATED badges below); the main view in the centre; an AI panel on the right, powered by HY4, with six actions: Morning Brief, Call prep, Notes from transcript, Update model, PM challenge, Export Excel. Each action first lists the inputs it will read and the objects it will produce, then runs step by step once confirmed: each step row shows a spinner that turns into a green check, the result is output as Markdown with clickable citations (news, transcript passages, model cells), the clock jumps to 08 / 10 / 11 / 13 / 16 / 17 in turn, and the produced files slide into the file tree. Market view: 4 macro indicator cards with sparklines, a 64-trading-day TSLA candlestick chart (MA20, event markers, OHLC on hover), a coverage heatmap, an AI-tagged news flow linked to thesis drivers, and thesis tracker cards (driver status On track/Watch/At risk, with a change log of reasons); at 07 a "0% APR" breaking headline slides in and driver #2 moves from Watch to At risk. Call view: the expert-call transcript on the left, with a waveform that fills as playback progresses; selecting a sentence lets you Cite it in one click as a bullet in the notes on the right, the source sentence is highlighted automatically, and hovering a bullet draws a glowing cyan-to-purple connector back to the source; bullets can be flagged model-relevant for Update model to read. Model view: an oversized target price and rating badge at the top (rating set relative to the coverage pool's average return), next to a tie-out table that reconciles line by line with the last exported Excel; below, grouped valuation inputs (market, WACC, SOTP, method weights) and key sliders (segment valuation multiples, FSD/Robotaxi/Optimus values, beta), bound two-way to the number inputs and the embedded Excel; further down, a SOTP breakdown chart, a yearly operating forecast table, HY vs consensus, bear/base/bull scenarios, DCF and reverse DCF, and a 7x7 sensitivity heatmap, all recalculated live on any change; changes count as unsaved until Save version, with version history; a rating flip triggers an alert and marks exports that depend on the old version as Stale. Outputs view: a timeline of exported Excel versions (each showing target price, rating, model version and generation time), with three tabs: Excel (sheet-by-sheet preview of the exported workbook), Dashboard (valuation football field, SOTP waterfall, revenue mix, margins and FCF, share price with a scenario fan), and Compare (pick any two versions and compare key numbers, SOTP changes, revenue paths and input differences). PM challenge turns the HY4 panel red and answers two challenge questions with calculations from the model's assumptions.\n\nAdditional Instructions: use plain HTML/CSS/JS only — no frameworks and no ES modules; double-clicking index.html must run it. All data is hard-coded sample data, and all interface text is in English. Colours: background #04060B with a faint 48px grid and two glows, one blue and one purple; cards are glassy with 14px rounded corners; accent colours #5B8CFF, #38E1FF, #9B6BFF, green #22D39A, red #FF5470, amber #FFB547; numbers in a monospace font. The business logic must be consistent throughout: the actions depend on each other (prep → notes → model → export → PM), and clicking a later step early first asks whether to run the missing prerequisite steps; the clock only moves forward; only one workflow can run at a time, with a Skip button to fast-forward; key driver values follow only the saved model version; an exported Excel is a snapshot of that moment and is marked stale when the TP deviates by more than 5% or the rating changes. The embedded Excel's Valuation sheet is the source of truth: the target price is the weighted SOTP, EV/Sales and DCF values, rounded. The footer reads: "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 header row and first column, sheet tabs at the bottom, a formula bar at the top; thousands separators, negatives in parentheses, percentages, multiples with an x; blue for inputs, black for formulas, green for cross-sheet references). 7 sheets in total: Cover, Drivers (revenue built by business line), IS, BS, CF, Valuation (WACC, yearly DCF plus terminal value, reverse DCF, SOTP, weighted target price, scenarios) and Estimate Changes; columns are FY2022–FY2025 actuals plus FY2026E–FY2032E forecasts, with over a thousand formulas, and the three statements tie every year (CF ending cash flows back into BS, and BS has a Balance check row for every year). Only input cells are editable; Enter recalculates the whole workbook and the changed cells flash; clicking any formula cell highlights, in numbered order, the reference chain from the inputs to that cell, with a side panel listing the value at each step; the header always shows the tie-out checks (balance sheet balances, cash ties, target price matches the Model view); the workbook can be exported in one click as an .xlsx that keeps formulas, formats and frozen panes, with identical values when opened in Excel. Implement the formula engine yourself in plain JS, without any library." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413073154_729df0f23d8001165a680a4704869a6a.jpeg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768854;2109128854&q-key-time=1793768854;2109128854&q-header-list=&q-url-param-list=&q-signature=3341872c3792ab78c5577c076a56856bd723f37c" [[items]] title="Iterative revisions of teaching presentation slides" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413092996_a329ccf6627806afc4bba102be7f9592.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768971;2109128971&q-key-time=1793768971;2109128971&q-header-list=&q-url-param-list=&q-signature=1276839aa6d4a65d1dc705b7179189800ea28ae0" description="Prompt: Round 1:\n\nI'm a 7th-grade science teacher. Please create a ready-to-teach deck, “How Do We Compare Paper Bridges Fairly.pptx” , and deliver only the editable PPT.\n\nDesign a 40-minute lesson that introduces new content. There are 32 students in 8 groups of 4, with no internet and no phones. Each group gets 9 sheets of 80 g/m² A4 paper of the same brand, two supports of equal height, a ruler, and 10 g weights. Experiments are done on the desk, with a tray underneath to catch anything that falls. No glass, and no loading with body weight.\n\nComplete core materials: in the reference experiment, each bridge is made from a single sheet, and a new sheet is used for every trial. Place the load at the center of the bridge and add 10 g at a time. Record the safe load, which is the last level before the first obvious collapse. If the bridge still hasn't collapsed at 120 g, record "≥120 g" and stop loading. Results from three trials:\n\nA, flat bridge, 15 cm span: 20 / 20 / 30 g\nB, single V-fold bridge, 15 cm span: 60 / 70 / 60 g\nC, corrugated bridge, 10 cm span: 90 / 100 / 90 g\nAll data are simulated.\n\nLearning objectives: students can clearly state which conditions are changed, which are measured, and which are kept the same; can propose a fair, repeatable comparison plan; and can draw conclusions that stay within the limits of the experimental conditions.\n\nHave students first make a prediction, then actually design or carry out the experiment, and then revise their prediction based on the evidence. Include an individual check for understanding, and give the time for each part of the lesson. Put teacher answers, common misconceptions, and follow-up questions in the speaker notes or on separate teacher-only slides, and don't show them before students respond. Diagrams, questions, and tables must be editable. You decide how many slides to use, and make sure everything is easy to read when projected. Don't treat data from different spans as enough evidence that one structure is better, and don't make up results as if this class had already measured them.\n\nRound 2 (AI editing on selected content):\n\n@tdoc-selection#1793678002089:"Slide 2 · Full lesson: 40 minutes" Put the timing on every slide, like this one, into the speaker notes. Only the teacher needs to see it.\n\n@tdoc-selection#1793678023214:"Slide 4 · The same sheet of 80 g/m² A4 paper of the same brand…" This text talks about three shapes, but the diagram below doesn't show any difference between them." poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413092572_576ecca397c25da99dcc09229979b5be.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793768969;2109128969&q-key-time=1793768969;2109128969&q-header-list=&q-url-param-list=&q-signature=76711c9803be480034e6c2603e9b78d211e6e5bb" [[items]] title="Business trip invoice reconciliation and financial audit" type="video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413110249_cff936180facb9019ea4c75b69344ac4.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793769063;2109129063&q-key-time=1793769063;2109129063&q-header-list=&q-url-param-list=&q-signature=27626c52636a2fa4564abb7c0f179a6ca0dae9ff" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413105949_b21b785631590bb85207ba13639acc0a.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793769062;2109129062&q-key-time=1793769062;2109129062&q-header-list=&q-url-param-list=&q-signature=28762a7adabbcad660e14c739d40badb5ef523d4" description="Prompt:Task 1: I'm Wen Bo from the R&D Department (employee ID E1312, level P4). From August 10 to 12, I traveled to Qingdao to set up and test equipment in a client's server room. The invoices from this trip are scattered across my QQ Mail inbox ([email protected]) and the current folder. Following company policy, please put all reimbursable receipts into a single zip file, along with an itemized expense summary. Flag any receipts that don't comply with the policy and explain why. Prepare everything to send to my colleague in Finance (WeCom: clarenceai). Sending the message with English.\n\nTask 2: \nPlease review all 24 expense claims in 03_Expense_Claims/ and decide whether each one should be Approved in full, Approved with deductions, or Returned for additional documents. You must read and cross-check all the materials in this directory yourself, including the policy documents and their appendices, the employee roster, department budgets, individual cumulative allowances, the historical invoice ledger, and the related emails. When you're done, give me your conclusions as a web page." @@@ @@@space height = 56 @@@ ## Scientific research and AI R&D With deep reasoning, scientific computing, and long-horizon execution, Hy4 pushes the frontier of end-to-end scientific research. On TenureTrail, our internal benchmark of 123 research tasks across seven disciplines, Hy4 ranks first among open-weight models. It shows strong capabilities in mathematical construction, physical modeling, earth sciences, systems engineering, and socioeconomic decision-making. We are also laying the groundwork for recursive self-improvement by advancing Hy4’s AI R&D capabilities. Paired with our Hyra 2.0 harness, Hy4 optimized our internal pre-training infrastructure for smaller models with the same architecture, and autonomously trained a variant that outperforms the official Jev baseline on multiple benchmarks. To make these AI-accelerated results reusable for the broader community, we are open-sourcing two foundation models co-developed by Hy4 and Hyra 2.0: HyCell for single-cell representation and HyPDE for physical field modeling. @@@mediaGallerySwiper [[items]] type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413161479_3931e8391916d00ebf64f11e5bf730b5.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793769378;2109129378&q-key-time=1793769378;2109129378&q-header-list=&q-url-param-list=&q-signature=9c45a59015edaf9688437eacb724eea46b1e4150" description="Hy4 Reconstructs a 3D Mouse Embryo from 7 Million Cells in 1.5 Hours on a Single Machine" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110400015543_23d03c33464f967b92db9b5e30c6ab1a.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793721718;2109081718&q-key-time=1793721718;2109081718&q-header-list=&q-url-param-list=&q-signature=142cdd5ead1ef3f712464eaeb6ed9708b28afb8c" [[items]] media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110402192830_964ebf3778ed952c8db86d57e1ad3431.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793729971;2109089971&q-key-time=1793729971;2109089971&q-header-list=&q-url-param-list=&q-signature=aa8d7d3cd398c3fd82c434f75f469e243ac411f6" description="[{type: 'text', content: 'In collaboration with a mathematics team at Peking University, Hy4 found a counterexample to the Méndez–Pinto–Villarreal conjecture, proposed new conditions to refine the conjecture, and formally verified the related proofs in Lean. This completed the full research workflow, from finding a counterexample to revising the statement and verifying the proof. Learn more:'},{type: 'link', content: 'https://github.com/frenzymath/SimisMonomial',url:'https://github.com/frenzymath/SimisMonomial'}]" [[items]] media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110402200438_8c613f6a2453c45131249da7cd1bc9a8.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793730007;2109090007&q-key-time=1793730007;2109090007&q-header-list=&q-url-param-list=&q-signature=3dc5c2f0163e17d7a1c9cd3f83bc0ef4e15058ae" description="Working on a smaller internal version of Hy4, Hy4 autonomously identified performance bottlenecks in our pre-training system and ran multiple rounds of infrastructure optimization across several areas, improving end-to-end training throughput by 20.1%." [[items]] media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110402202743_19dff19e882464abfd963f70a8ec7755.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793730030;2109090030&q-key-time=1793730030;2109090030&q-header-list=&q-url-param-list=&q-signature=614d9dc28abdbdc1d17414ab720e1ad2cd42d8c0" description="Paired with Hyra 2.0, Hy4 automatically trained a Jev model. The resulting Hyra-Jev outperforms the official Jev on 6 of 8 benchmarks." [[items]] media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110402212194_218211c5f701f1190bccfb9f6deed1ce.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793730084;2109090084&q-key-time=1793730084;2109090084&q-header-list=&q-url-param-list=&q-signature=2fe461dca5bc144c861f5f9533b4dd0d9c02ee85" description="Using Hyra 2.0, Hy4 developed HyCell, a foundation model for single-cell representation. On seven cell atlases held out from training, HyCell matches the overall performance of Stack, the current state-of-the-art model, with just 16 million parameters, about one-fourteenth as many. It also achieves the highest average biological fidelity score among the models evaluated." [[items]] media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110402213716_05c531fa72d7a95dfc01c529ee1d739c.jpg?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793730100;2109090100&q-key-time=1793730100;2109090100&q-header-list=&q-url-param-list=&q-signature=8d0e051656c3563d9bed160c1ccb5d36bd9664f3" description="For physical field modeling, Hy4 worked with Hyra 2.0 to develop HyPDE, a physics foundation model. Compared with Walrus, the current state-of-the-art model, HyPDE achieves lower prediction error on 9, 10, and 8 of 13 physics simulation datasets for single-step, short-term (1–20 steps), and long-term (21–60 steps) prediction, respectively." @@@ @@@space height = 56 @@@ ## Long-horizon personal assistant With stronger context learning capabilities, Hy4 can synthesize complex information from multiple sources. Through WorkBuddy’s integration with WeCom, Hy4 connects directly to everyday work context, including chat history, emails, documents, calendars, and meeting notes. When you’re managing multiple projects at once, Hy4 can track each project’s history and latest updates, surface action items, prepare meeting briefs, flag upcoming deadlines, and help prioritize tasks. Hy4-powered team collaboration features will also launch in WorkBuddy alongside the new version of Projects. @@@mediaGallerySwiper [[items]] title="Bring together what needs my attention across multiple sources" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413224018_71f83641a27f1d00cf7da54875cbbd6d.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793769760;2109129760&q-key-time=1793769760;2109129760&q-header-list=&q-url-param-list=&q-signature=ab056b5ab008c68714d0e971bdcec700721189e8" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413223759_fafd01aac799e6355dfb28b045fce72c.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793769759;2109129759&q-key-time=1793769759;2109129759&q-header-list=&q-url-param-list=&q-signature=6e6fef6cb0ca1f472346a38a58bd70f6b12ca83a" description="Prompt: Goal: Look at my chats (group chats, private chats and files), emails, calendar, meetings and documents. Use them to make a "Daily Focus" board of my tasks for each day. Give it to me as one HTML file that I can click around in and open without the internet. Don't just sum up my chats or list everything that happened. I want to see right away what matters today and what I should do first.\n\nContent: Pull the newest information from all of these places and remove anything that shows up twice. Find the things that matter in my work, my personal life and my daily chores. Sort them by how important and how urgent they are. For each one, say clearly what happened, where it stands now and what to do next. Write for anyone, in simple, everyday words. Don't assume the reader knows the background. Avoid jargon, inside terms, short forms, hard-to-follow sentences and the stock phrases AI tends to use. Keep the sources that matter." [[items]] title="Analyze my fitness history using records from the past several years" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413245195_90bc75db54c33a7f72c14ee71e66365a.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793769897;2109129897&q-key-time=1793769897;2109129897&q-header-list=&q-url-param-list=&q-signature=3a2d79d4c2f704eac8ad0a984e4741f4b24ac4d3" description="Prompt: I haven't worked out in a long time, and I want to start again. In 2024, I stopped making progress. Can you help me find out why? Was it how I trained, the supplements I took, or my sleep? And do you have any tips for getting back into working out?" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413244849_6d977c6cc2885e481cd0b5b1da3a11eb.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793769891;2109129891&q-key-time=1793769891;2109129891&q-header-list=&q-url-param-list=&q-signature=0e549c8398920727f24eeeff8d85dc795ff5733d" [[items]] title="Long-term collaboration with a personal assistant" type = "video" media_src = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413264827_c7afb7b2b51dac43544dbd1c330f9cf7.mp4?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793770011;2109130011&q-key-time=1793770011;2109130011&q-header-list=&q-url-param-list=&q-signature=b7656b361fa27c82ed696564ebe48fa255f35d88" poster="https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/a951c950/2026110413264495_d6668a3034f8d09da4d78d23603472d3.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793770008;2109130008&q-key-time=1793770008;2109130008&q-header-list=&q-url-param-list=&q-signature=c6c9c6740ad382bb3b6f7612a8bb7771dccc339e" description="Prompt: 1. I'm starting my internship next week and meeting my mentor for the first time, so I'd like to give them a clear picture of my background up front. Can you help me figure out how to introduce myself, covering what I've worked on, what I know well, and what I'm hoping to learn from the team? Keep it natural, though. I don't want it to sound like I'm reading off my resume.\n2. My thesis is finally done. Let's wrap things up by writing the acknowledgments together. I'd like to open by thanking everyone who helped me along the way. Then I want to talk about my research in grad school and during my internship, how my focus shifted, and what I learned." @@@ @@@space height = 56 @@@ # Product feedback @@@testimonialSlider ariaLabel = "产品反馈" itemsPerPage = 2 mobileItemsPerPage = 1 loop = false defaultActivePage = 0 showCount = true countSeparator = " / " countMinDigits = 2 countLabelTemplate = "第 {current} 页,共 {total} 页" emptyText = "暂无证言内容" contentEmptyText = "暂无证言内容" previousLabel = "上一页证言" nextLabel = "下一页证言" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110401453515_2a709c2caeeff98490a465548f2c17e0.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793727938;2109087938&q-key-time=1793727938;2109087938&q-header-list=&q-url-param-list=&q-signature=9e0a69e3fd0992032e69faa0db5f5d3dfcc22ce3" brandName = "Manus " content = "In our internal evaluations, Hy4 delivered strong results across deep research, data analysis, and office document tasks, ranking among the top models overall. It performed particularly well in document creation and editing, achieving a task completion rate of nearly 97%. Hy4 also demonstrated strong instruction-following capabilities, accurately handling complex requirements such as citation formatting, heading numbering, and detailed revision instructions. For Manus users, this means less rework, faster execution, and more reliable results in real-world workflows." authorName = "Manus " [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403103430_6c82b854d9c3e7d94c315d33d9e663a2.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793733036;2109093036&q-key-time=1793733036;2109093036&q-header-list=&q-url-param-list=&q-signature=c1a94610a3b335c4b0ed06f5fd2128cfaf507af3" brandName = "RabyteTech" content = "As an AI service provider focused on investment research, we evaluated Hy4 on the IRAB benchmark for investment research agents, using 101 investment research tasks. Hy4 scored 61.63 overall, 12.03 points higher than Hy4 preview, showing clear progress among the models that we’ve been tracking. Hy4 is firmly in the top tier of Chinese models and has further narrowed the gap with leading proprietary models. The value of an investment research agent lies in delivering outputs that are truly usable, and Hy4 is improving quickly at completing investment research tasks end to end. Hy3 has already been integrated into some AlphaPai products, and we look forward to deepening our collaboration with the Hy team to push the frontier of AI in finance." authorName = "RabyteTech" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403101337_83bb66953638a161a2eb8dccc366d1e2.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793733015;2109093015&q-key-time=1793733015;2109093015&q-header-list=&q-url-param-list=&q-signature=fefdcaa36f2e8137e2964971559db3f2c7d46536" brandName = "Comein" content = "As the team behind a leading AI investment research workbench for financial institutions, we evaluated Hy4 across our internal investment research agent tasks. Hy4 is currently the only Chinese model to tie with Claude Opus 5.5 for second place overall. On data queries, it scored 75.4—nearly 20 points ahead of GPT-6.1 Sol—ranking first among all models evaluated. In investment research, getting the data right is step one: if the underlying data isn't accurate, none of the analysis that follows matters." authorName = "Comein" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403111834_c9be733d18d77fde0da2befde1805315.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793733080;2109093080&q-key-time=1793733080;2109093080&q-header-list=&q-url-param-list=&q-signature=dbc08d21602b3fb0d9c913c0c3a56e641d9e4974" brandName = "Caidazi (财搭子)" content = "Paired with our Caidazi Harness framework, Hy4 delivers strong value in investment research and decision support. In our evaluations, it ranked #1 in single-stock research (4.27/5)—scoring highest (including ties) across five models on 42.9% of responses—while achieving 4.43/5 in deep research and 4.03/5 in ongoing monitoring. Hy4 structures analysis around core drivers and uncovers strong causality behind weak correlations: tracking how orders convert to profit and cash flow in company research; breaking down what it takes for a sector thesis to play out; separating index performance from premium risk in ETF analysis; building scenario plans around cost basis for decision support; and turning key price levels and risk signals into watch triggers. It translates professional analysis into clear decision criteria—complete with risk boundaries and invalidation signals—supporting investors from initial thesis to ongoing tracking." authorName = "Caidazi (财搭子)" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403105674_c8393dd0108b1de1b6734a2b43233d13.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793733058;2109093058&q-key-time=1793733058;2109093058&q-header-list=&q-url-param-list=&q-signature=bcd06d9e4c6b20233a21e1a41b1375d16b5e92ff" brandName = "China Southern Asset Management" content = "Financial services need to address investors’ specific questions about markets, industries, and financial planning. We ran an internal evaluation of Hy4 across 398 questions on market outlooks and investment planning. When asked about the factors driving the nonferrous metals sector, for example, Hy4 breaks the analysis down across six dimensions—macroeconomic conditions, supply and demand, policy, costs, financial markets, and metal-specific properties—while clearly distinguishing the drivers behind industrial versus precious metals. This makes Hy4 well-suited for investor education and financial Q&A, helping investors build a systematic understanding of the market." authorName = "China Southern Asset Management" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403114944_e4d8a1908969c3b3918276b2f46057e2.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793733111;2109093111&q-key-time=1793733111;2109093111&q-header-list=&q-url-param-list=&q-signature=4dde16420f09abc7ad256ee35f2bf58b10839497" brandName = "China Universal Asset Management" content = "Investment research demands strict verifiability—especially across macroeconomic views, sector fundamentals, and valuation assumptions. We tested Hy4 on our in-house research agent platform across core workflows like macro event analysis, valuation modeling, portfolio construction, and risk assessment. Hy4 accurately grasps user intent and explicitly lays out its calculation logic and underlying assumptions. Because researchers can audit how every conclusion is reached and judge whether it fits their specific thesis, the model’s outputs are far more trustworthy." authorName = "China Universal Asset Management" @@@ @@@space height = 40 @@@ @@@testimonialSlider ariaLabel = "客户证言" itemsPerPage = 2 mobileItemsPerPage = 1 loop = false defaultActivePage = 0 showCount = true countSeparator = " / " countMinDigits = 2 countLabelTemplate = "第 {current} 页,共 {total} 页" emptyText = "暂无证言内容" contentEmptyText = "暂无证言内容" previousLabel = "上一页证言" nextLabel = "下一页证言" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403005137_38f7d1fa6bb1f062c21920b162d05167.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732454;2109092454&q-key-time=1793732454;2109092454&q-header-list=&q-url-param-list=&q-signature=0a456a5b194345dad301414e230221ee224e7ae9" brandName = "WorkBuddy" content = "Across web tasks, visual creation, and 3D use cases, Hy4 strikes a much better balance between quality and speed. In multi-step browser tasks, it uses feedback from its actions to adjust its approach and verify results on its own. In our evaluations, Hy4 achieved an average score of 83.18 and a 96.15% pass rate, ranking #1 among Chinese models in 9 of 12 scenarios." authorName = "WorkBuddy" authorTitle = "Anne, Product Manager" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403012828_b52b5e93c4dadeea0b56cb690a633d10.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732490;2109092490&q-key-time=1793732490;2109092490&q-header-list=&q-url-param-list=&q-signature=c02d00c5985f6709fcf074128d697a7b38746961" brandName = "Yuanbao" content = "On our internal evaluation set, Hy4 scored 72.53, up from 65.25 for Hy4 preview, with particularly strong gains in multi-turn conversation and writing. Multi-turn memory improved by 6.67 points, bringing better reference resolution and smoother flow across turns, while writing quality jumped 10.89 points with much greater content diversity. Compared with our previous VLM, Hy4 also boosted multimodal agent task completion by 36.8% and multimodal Q&A by 15.8%, delivering major gains in image understanding, joint image-text reasoning, and cross-modal context." authorName = "Yuanbao" authorTitle = "Nyx and Ranmer, Product Managers" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403015619_135ff144dd6a08bc8d026b733198b18b.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732518;2109092518&q-key-time=1793732518;2109092518&q-header-list=&q-url-param-list=&q-signature=af05b846c612b6683f88d2bfcc794c8daac110e8" brandName = "Honor of Kings" content = "On Honor of Kings, Hy4 brought significant upgrades to both our development and testing workflows. When paired with our automated platform on 20 game configuration tasks, Hy4 improved accuracy from 91.74% to 97.28%, reaching 99.32% after the agent’s own verification and self-repair. On our automated bug analysis platform, code root-cause analysis accuracy jumped from 59.5% to 73.8%. We also saw 3.7x greater depth in tool calls than with Hy4 preview, with no noticeable increase in time or token usage. Together, these improvements gave us an across-the-board upgrade in AI game production." authorName = "Honor of Kings" authorTitle = "Jesse, AI Gameplay Design Lead" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403022051_03369c2083d0b60333200772f1fa4f5f.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732542;2109092542&q-key-time=1793732542;2109092542&q-header-list=&q-url-param-list=&q-signature=a61c2879c95a0ecbdd9e314431ad85465c98159f" brandName = "Peacekeeper Elite" content = "Hy4 made a strong impression in our AI NPC evaluations for Peacekeeper Elite. Compared with Hy4 preview, its dialogue feels much closer to natural squad banter, with more lifelike roleplay and stronger persona consistency. It’s also far more proactive—breaking the ice on its own, picking up naturally on player cues, and keeping conversations flowing. By offering timely encouragement, validation, and follow-up questions, it builds a genuine sense of connection. On a 3-point human rating scale, Hy4 averaged 2.91 overall and hit a near-perfect 2.96 on emotional support—making AI NPCs feel like real squadmates fighting alongside you, just like playing with friends." authorName = "Peacekeeper Elite" authorTitle = "ELI, Project Management Lead" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403023983_8ca88aeaa8e8a39d2774d600860162f9.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732561;2109092561&q-key-time=1793732561;2109092561&q-header-list=&q-url-param-list=&q-signature=29ee4e6413e3333f9e0354abb9a212bc37d68856" brandName = "QQ" content = "On QQ, Xiao Q (QQ AI Agent) helps users look up information, set reminders, summarize unread messages, and brainstorm social ideas every day across more than 40 scenarios. Compared with Hy4 preview, Hy4 increased overall user satisfaction from 90.11% to 93.10%. In three high-frequency scenarios—web search, Qzone search and recommendations, and subscription reminders—satisfaction improved by 6.66, 4.69, and 2.52 percentage points, respectively." authorName = "QQ" authorTitle = "Kedeng, Product Manager" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403051259_3cce84d5fe7d7f0402c18f161e606bc8.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732714;2109092714&q-key-time=1793732714;2109092714&q-header-list=&q-url-param-list=&q-signature=dd2571bdce3195823e3d88b9cf0532124c9e4a98" brandName = "WeCom" content = "In WeCom Smart Assistant’s workplace collaboration scenarios, Hy4 can break down instructions with multiple intents into subtasks and independently plan tool calls. Compared with Hy4 preview, we saw improvements of 13.5% in instruction following, 12.7% in complex information retrieval, and 8.5% in overall chat performance." authorName = "WeCom" authorTitle = "Dylan, Algorithm Lead" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403053684_6db909f2a73575b787b375e565e7734c.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732738;2109092738&q-key-time=1793732738;2109092738&q-header-list=&q-url-param-list=&q-signature=68f70aba33466536ef2c983dc0b03c2083077604" brandName = "WeRead" content = "In our evaluations at WeRead, Hy4’s overall score rose to 87.69 from 82.10 for Hy4 preview. The most noticeable improvements were in how it prioritizes information and controls length—up 22.2% and 50.7%, respectively. In tasks such as mapping character relationships, interpreting plots, and answering knowledge questions, Hy4 more accurately zeroes in on what each question is asking and cuts unnecessary, redundant content. Answer clarity and adaptation to different readers also improved by 3.0% and 6.6%, respectively, creating a better interactive reading experience overall." authorName = "WeRead" authorTitle = "Isaac, Algorithm Lead" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403045256_0c1336139dd39776eaf99aa0854625ee.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732695;2109092695&q-key-time=1793732695;2109092695&q-header-list=&q-url-param-list=&q-signature=92a388d0039937338500fc78fabc67566fc0d4c5" brandName = "Tencent News" content = "For news search and open-ended Q&A, we ran blind evaluations using a test set of real user questions. Compared with Hy4 preview, Hy4 improved by over 28% across accuracy, level of detail, and fluency, driving a 36% increase in overall end-to-end answer quality." authorName = "Tencent News" authorTitle = "Aero, Product Manager" [[items]] logo = "https://hy-model-ap-prod-1258344703.cos.ap-guangzhou.myqcloud.com/llm-blog/default/9a5ea02f/2026110403035813_ba938573b46ba47b93f4d3805db90d82.png?q-sign-algorithm=sha1&q-ak=AKIDRl074nOsGdJ9zjMsCRWP3ShmgS3VtX4S&q-sign-time=1793732640;2109092640&q-key-time=1793732640;2109092640&q-header-list=&q-url-param-list=&q-signature=98a64ee10ab7bd7d0c701c6c22345ec3b0aa8564" brandName = "ima" content = "In our evaluations across ima’s knowledge-base, general Q&A, and agent scenarios, Hy4 delivered improvements of 9.7% in response quality, 3.4% in reasoning accuracy, and 1.2% in overall chain-of-thought performance compared with Hy4 preview. Its responses get to the point more effectively and are backed by stronger reasoning. " authorName = "ima" authorTitle = "Kaycee, Product Manager" @@@ # Responsible open-sourcing and release We recognize the potential safety risks posed by advanced cyber and biological capabilities, and we are committed to a more responsible approach to model training, open-sourcing, and release. During Hy4’s training, we limited cybersecurity data to defensive material related to proof-of-concept (PoC) generation and vulnerability patching. We filtered out offensive content such as exploit development, reverse engineering, and penetration testing to avoid amplifying potential offensive capabilities. We also worked with experts from Tencent Xuanwu Lab and Tencent Keen Security Lab to conduct red-team audits covering cryptanalysis, reverse engineering, and vulnerability exploitation. Their findings indicate that Hy4 does not pose elevated risks in offensive tasks such as reversing cryptographic algorithms, defeating binary obfuscation, or exploiting vulnerabilities. In biosecurity evaluations, Hy4’s open-ended experimental design capabilities remain well below our high-risk threshold, with a score of 30.4/100 on Hy-BioExperimentBench. Internal experts audited the model’s reasoning traces and found that although Hy4 has a solid grasp of foundational biology, its ability to turn that knowledge into actionable experimental protocols, a critical bottleneck, remains limited. We found no evidence that Hy4 meaningfully lowers the barrier to causing biological harm. Hy4 is now available on Tencent Cloud and OpenRouter, as well as in products including CodeBuddy, WorkBuddy, Yuanbao, and ima. It delivers major performance gains over Hy4 preview at the same API price. @@@space height = 32 @@@ ::: center Hy4 API pricing (per million tokens) ::: ::: center |Cached input | Input | Output | | ------ | ------ | ------ | | $0.042 | $0.834| $2.501 | ::: @@@space height = 32 @@@ We plan to open-source Hy4’s model weights and have partnered with the vLLM (Inferact Inc.) and SGLang communities to provide ecosystem support. When we release the weights, we will also publish a comprehensive system card detailing our safety evaluations and additional technical details. # Looking ahead Although Hy4 is one of the strongest open-weight models available today, a notable gap remains between it and top proprietary models, particularly on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering. We will continue to steadily scale pre-training and post-training to narrow this gap.
We also plan to release Hy4 flash in the near future, with significantly improved cost-efficiency, faster inference, and closer co-design with products to improve utility. # Appendix: benchmark results 