Hunyuan Hy4 Official Release: A Cheaper, Multimodal Agent Model That Can Drive Another AI

Hi everyone, I'm Brother Er.

Tencent's Hunyuan Hy4 official version is out, and I immediately started using it in WorkBuddy.

Believe me, this model is really strong.

Let me record the screen so you can feel it. This is what I got by running WorkBuddy + Hunyuan Hy4 official.

Vibe knowledge showcase of Ohm's law, isn't it nice?

Compared to the preview version, the Hy4 official version expanded pre-training and post-training scale, introduced native multimodal visual capabilities and Computer Use capability for the first time, maintained the same API price as preview, and plans to open-source model weights.

According to the official benchmark results, the Hy4 official version has improvements across multiple dimensions. In the Agentic Coding direction, Terminal-Bench 4.0 went from 26.3 in preview to 44.4, WebCraftBench, which tests webpage generation, went from 52.7 to 78.9, 3DArena-Modeling went from 23.4 to 37.1, and GameDev and GameLogic benchmarks, polished jointly with Tencent IEG, improved by 19.1% and 9.8% respectively.

From Agent Arena evaluation data, Hy4 achieves a good balance between output and cost, sitting on the "Pareto frontier." Look at the chart below: the vertical axis is the evaluation score, the higher the better; the horizontal axis is the median per-task cost, the further to the right, the cheaper. Based on more than 8,500 real long-horizon agent tasks from global users, Hy4 costs $0.55 per task, about 62% lower than Kimi K3.1 Pro's $1.46.

Of course, just looking at scores isn't intuitive. Follow me and enjoy the capabilities of Hunyuan Hy4 official. I've organized all the prompts so you can reproduce them.

01, Vibe knowledge showcase: Ohm's law

I've always had this idea: use Vibe format to re-record those obscure knowledge points from middle and high school, including math, physics, chemistry, and biology.

Just in time, Hunyuan Hy4 official was released, and with WorkBuddy, I can make this idea real.

This is a one-click production of Ohm's law. I'm sure you'll give it a thumbs up, 😄

First, let's list the knowledge points.

You can see WorkBuddy nicely organized a markdown document for me, and it's quite comprehensive.

This time I chose Ohm's law as the topic. First, I asked WorkBuddy to soft-link my previous Skill. The advantage is that the Skill only needs to be maintained once, but can be used by multiple Agents.

Then we input the prompt.

Use the one-minute video Skill to make Ohm's law, first produce the script.

WorkBuddy is quite playful, "Don't worry, I'll handle it," and during the process it said, "It's trying hard to look busy (actually really working)." I find it adorable.

Honestly, when talking to AI every day, the worst thing is the "AI smell." Hunyuan Hy4 this time is full of human touch. It feels like the person across from me is no longer a cold AI, but a Personal Agent that's great at providing emotional value.

Soon the script was ready, and the quality felt high.

You can see it was very rigorous, even looking up "Ohm's Law Lesson Plan," the "Resistance" entry, and "2022 Junior High Physics: Ohm's Law and Applicability Conditions."

The quality of the script determines the quality of the video, so this step is very important.

I made a few minor adjustments to wording and phrasing, then synthesized the audio and listened to it. Once it was fine, we moved to the next step.

Connect a fixed resistor and an ammeter in series, then attach one battery; the ammeter needle deflects a bit. Replace with two batteries, and the needle deflects more. The resistance hasn't changed, the voltage has increased, and the current has also increased.

Current is the amount of charge passing through a conductor's cross-section per unit time. Voltage is the cause of current. Resistance represents the conductor's opposition to current. These three quantities must correspond to the same conductor.

Ohm's law says: the current through a conductor is directly proportional to the voltage across it and inversely proportional to its resistance. Voltage is denoted U, in volts; current is I, in amperes; resistance is R, in ohms. The formula is I = U / R.

Use simple integers as an example. Suppose the fixed resistor is 3 ohms, and 6 volts across it; current equals 6 divided by 3, so 2 amperes. If voltage rises to 12 volts, current becomes 4 amperes. Keep voltage at 12 volts, change the resistor to 6 ohms, and current returns to 2 amperes.

Ohm's law is an empirical law derived from experiments. Take the same fixed resistor, keep temperature roughly constant, add batteries one by one, and record voltage and current each time: when voltage becomes several times larger, current also becomes that many times larger. Keep the voltage across the resistor constant, make the resistance several times larger, and current becomes a fraction of the original. Within the error margin, each set of measured data satisfies I = U / R.

In summary: when resistance is constant, the larger the voltage, the larger the current; when voltage is constant, the larger the resistance, the smaller the current. So current can be calculated by dividing voltage by resistance. Ohm's law applies to metal conductors whose temperature remains roughly constant; gas conduction and components like diodes do not obey this law. Knowing any two of voltage, current, and resistance lets you calculate the third, such as determining what resistor to put in a circuit based on supply voltage and required current.

Once finalized, WorkBuddy called a TTS script to synthesize a 97-second voiceover. I listened and it was fine.

The storyboard came too. In OUTLINE.md, WorkBuddy divided the entire video into 13 semantic groups and 11 Memphis-style scenes, ensuring the whole video shares the same circuit components.

Next step, start producing the first chapter. To allow timely corrections, I kept each chapter to about 30 seconds.

The first chapter came out. Overall it looks great, visually pleasing. No random generation at all, it got it in one take. I have to say Hy4 really knows how to direct video production.

The only flaw: with one battery, it shouldn't reserve a spot for a second battery, otherwise the circuit won't be closed.

I sent the screenshot to WorkBuddy directly to test Hy4's multimodal understanding.

OK, it fixed it.

WorkBuddy recognized the light purple empty battery slot and broken wire in the screenshot, changed the one-battery state to a complete closed loop, and then completed the series connection when switching to two batteries.

Next, produce the other chapters and start the Remotion Studio preview service on port 3123.

02, Sheep grazing mini-game

The second case tests pure code drawing and game generation. Without any sprites or external image assets, I had WorkBuddy drive Hy4 official to draw a watercolor-texture sheep farm using native Canvas 2D.

Prompt as follows.

Make a complete "healing sheep grazing" web mini-game, and deliver directly runnable source code.

[Visual reference]
Visual points: cream background, slight top-down view of a miniature grassland, soft textured light green grass; a winding blue stream on the left, a cream dirt path on the right-center; irregular dark green field ridges dividing the grass; green trees with rounded foliage and one pink-purple tree; ten fluffy ivory-white sheep with dark faces, short legs, and soft shadows on the ground; slowly drifting clouds in the sky, small yellow flowers scattered on the grass. Overall refined, soft, with a watercolor and miniature model feel. If the link doesn't open, follow the description above.

[Technical requirements]
1. Use real Canvas 2D or Three.js to programmatically draw all visuals. It is forbidden to use reference screenshots as backgrounds or to use image assets.
2. Single page, responsive on desktop and mobile, scene should fill the viewport as much as possible; interface in minimalist Chinese.
3. Dependencies written into package.json with locked versions, include a README explaining local startup.

[Simulation and behavior]
1. Sheep autonomously find grass and lower their heads to eat; grass gets consumed and slowly regenerates over time.
2. The flock maintains spacing, avoids streams, dirt paths, trees, and grass boundaries, and does not walk into water.
3. Day/night toggle: at night the light turns cold, sheep move slower, grass grows slower, and you can add moonlight, stars, and fireflies.

[Interaction]
1. Click or touch the grass: scatter a circle of fresh grass, attracting nearby sheep to come over.
2. Click or touch a sheep: it jumps and a heart appears.
3. Buttons: add a sheep, remove a sheep, pause/resume, reset, day/night; UI shows sheep count and grass lushness.
4. Support keyboard shortcuts (e.g., Space to pause, D to toggle day/night), and explain them in the UI or README.

[Deliverables]
Complete source, README, a runnable local preview URL, and real-browser interaction tests (click, touch, mobile viewport, no console errors) with test screenshots. Put all files in a single new project directory, do not modify other projects, do not commit, push, or deploy.

I played with it. It's very healing.

Clicking the grass makes the sheep run over quickly.

Because clicking the grass is like planting grass, so all the sheep come over to eat.

At night, there are fireflies. Wow, that detail is fantastic.

Clicking a sheep gives a heart 💗, but it's hard to screenshot because the action is fast. You'll feel it from the recording.

03, Hand-drawn style "How to make milk tea"

The third case has stricter limits: no Skills, no frontend frameworks, chart libraries, or animation libraries. Pure HTML, CSS, vanilla JavaScript, and SVG to hand-draw a dynamic "How to make milk tea" flowchart.

The prompt is as follows.

Make a Chinese hand-drawn dynamic flowchart webpage "How to make milk tea" using pure JavaScript. Do not use any Skill, no frameworks, chart libraries, animation libraries, or image generation tools. HTML, CSS, vanilla JavaScript plus SVG/Canvas are enough; graphics are drawn by code, and opening index.html directly runs it.

Refer to the original: cream paper, cobalt blue crayon line sketch, slight paper texture and colored pencil hatching, grass-green/tomato-red/warm-yellow accents, large title and hand-drawn annotations, characters and utensils interspersed beside the flow. Keep the refined and fun dynamic state diagram feel, but change the subject to milk tea.

Three coherent chapters: 1. Good tea base (tea leaves into pot, hot water extraction, filtering); 2. Pearls turn Q-elastic (water boils, pearls in pot, cook through, simmer, mix with syrup, can run in parallel with steeping tea); 3. A cup is born (pearls into cup, add tea, add milk, sugar/ice choice, shake, seal and serve). Use real motion like steam, tea color changes, pearl tumbling, pouring liquid, milk-tea mixing, cup shaking to explain. The flow lines and animation must match the process. An unfinished previous step cannot jump to the finished product.

The whole graphic must have large, carefully drawn characters, teapot, strainer, pot, measuring cup, shaker, sealing machine, and finished product. No emoji or a few text boxes. The flow has parallel merges and visible rework branches: tea too strong -> add water to adjust; pearls undercooked -> continue simmering. Short, easy-to-understand Chinese. At the bottom: play/pause, previous/next, replay, speed, step progress; support Space and left/right arrows; sugar and ice options that change the finished drink. Rewind, pause, replay states accurate. Desktop-first, mobile can fully read controls.

Create a new standalone project, do not read or reuse other Case outputs. Implement directly, start preview, self-test, give a URL. Save this prompt as PROMPT.md, record the model display name, run method, and keep a snapshot of the first working version before iterating. Do not commit or push.

The resulting output is high quality.

Every step is well executed.

Looking at the project directory, WorkBuddy put all logic into three files: index.html, styles.css, and app.js.

index.html provides the page skeleton and bottom control bar; styles.css implements the cream paper and cobalt blue line sketch; app.js uses native SVG to draw the teapot, steeping, shaker, and two operating characters. When choosing different sweetness levels, the syrup height in the finished cup, the number of ice cubes, and the icons on the recipe card all change synchronously.

04, Building an ESP32 interactive learning website

The fourth case tests 3D hardware modeling and circuit simulation. When learning embedded development, if you don't have a dev board, static wiring diagrams can easily confuse you. This time I had WorkBuddy drive Hy4 official to make an ESP32 interactive learning website.

Prompt as follows.

Turn complex hardware into simple, fun lessons: make a Chinese-language ESP32 interactive learning website.

First complete a playable full version: a step-by-step course navigation on the left, a rotatable, zoomable, resettable 3D ESP32/breadboard/wiring scene in the center, and on the right brief theory, wiring instructions, and corresponding code. Below: run/stop, reset, and serial output. Include at least four lessons: light up an LED, button control, PWM dimming, and a mini piano. Switching lessons must really switch circuits, code, and interactions. The mini piano must have eight distinct notes, support mouse and keyboard A S D F G H J K, with keys, pins, sound, and serial output synchronized; provide auto-play and volume control. Browser simulation must be clearly labeled, not pretending to connect to real hardware. Choose a specific ESP32 model and use consistent pins and code.

Visual approach close to reference: clean light colors, professional but fun, with a fully operable hardware scene taking the main space; components have detail, annotations are clear, no overlap or clipping. Works on desktop and mobile. Do not just deliver a static homepage or feature description.

Create a new standalone project in hy4-esp32-course-run-01, do not read or reuse other Case outputs from the repository. Implement directly, start a local service and self-test real interactions, give a browsable preview URL. Save this raw prompt as PROMPT.md, record the actual model display name and run method; keep a snapshot of the first working version before fixing. Do not commit or push.

The result is also excellent.

You can rotate it with the mouse.

Click the "run program" button and you can really see current.

It even makes me want to buy a real board and hook it up.

05, Computer Use control

We already made the Ohm's law video preview, right, but it hasn't been rendered.

Next, we use WorkBuddy + Hy4's Computer Use ability to control a DeepSeek Harness to render the video.

Usually, when I'm away from the computer, I use this method: let one Harness control another Harness to get things done.

The advantage is that the controller is diligent, and the controlled one doesn't fall into a self-absorbed state. Using one Harness to supervise another usually produces far better deliverables than a single Harness.

You can try it.

If you're not sure whether WorkBuddy has Computer Use, you can enter this prompt.

Use computer operation, switch to DeepSeek Harness, read the last reply in the current chat and tell me. Don't type or send anything yet.

Usually the first time you need to enable some system settings, like privacy and security.

If you can get the content from DeepSeek Harness, it means we can use WorkBuddy to control DSH.

Next, we wait for DeepSeek Harness to render.

OK, WorkBuddy also returned.

Nice.

ending

It's visible that domestic models' capabilities are improving, and by a considerable margin. Agent Arena's evaluation data shows Hy4 ranked 14th on the overall leaderboard, tied with Kimi K3.1 Pro as the top open-source model.

Also, evaluation data from the large-model evaluation agency Artificial Analysis shows Tencent Hunyuan Hy4 scored 35, tied with Grok 4.7, and is second among domestic models after Kimi K3.1 Pro.

Hunyuan Hy4 surprised me in that no matter the type of case, the completion quality is high.

What surprised me even more is that in every conversation, you feel a living person in front of you. Little details like "Coming right up, faster than delivery" - I don't know which product manager designed that, but I really like it.

It has a human touch.

Actually, after dealing with AI for a long time, we need these little surprises; otherwise we're just dryly waiting for model output and tool calls, which is really boring.

Finally, here are the six core capabilities of the Hunyuan Hy4 official version.

See you next time.