Tencent Hunyuan Hy4 Debuts with Native Vision and Computer Control

Wenle, reporting from the Non-Famous Temple Quantum位 | WeChat Official Account QbitAI

Hunyuan Hy4 has officially handed in its exam paper!

Tencent isn't holding back this time — its attitude is written plainly on its face: join the ranks of the strongest open-source models.

770B total parameters, 49B activated parameters, the first introduction of native visual capabilities, and the first support for computer operation.

Instruction following, long-document understanding, and long-horizon Agent capabilities have all been upgraded together, with computer operation and long-horizon Agent tasks squeezing into the front ranks of the open-source camp.

From preview to the official release, Hy4 significantly expanded the data and training scale for pre-training and post-training, and even went straight to first-line businesses to engage in "Co-Design" joint polishing.

The domestic and international versions of WorkBuddy have already completed first-batch integration simultaneously!

So I first gave it a tough challenge in WorkBuddy: whip me up a fully interactive 3D mechanical keyboard.

75% layout, 84 keys, with keycaps, switches, positioning plate, PCB, battery, and casing — all six layers of structure accounted for.

You can drag it with the mouse to explode it in place and inspect the internal build quality, put it back together and type on it directly, and it even factors in the physical animation of key presses and spring-loaded rebound.

Looking at it this way, doesn't someone else now handle the job of making new product showcase demos? Right, Hy4～

Hy4 is already live on Tencent Cloud TokenHub, WorkBuddy/CodeBuddy, Yuanbao, WeCom intelligent assistant Dayuan, ima, and other products; the API is simultaneously integrated into platforms such as OpenRouter, Hermes, and OpenCode.

Tencent ecosystem products such as QQ, Honor of Kings, Peacekeeper Elite, WeChat Read, and Tencent News are also continuously integrating it.

The model weights will also be open-sourced in the near future.

So without further ado, let's run a round of hands-on testing first.

Pixel-Level Visual Perception

With the visual understanding introduced for the first time in Hy4, detail recognition is already fine enough to locate roughly 2-meter rooftop components in a 2.8-billion-pixel aerial photograph.

That eyesight is genuinely impressive.

So next, let's test its eye for detail.

First task: follow a design screenshot and build a retro-style music player.

Never mind high-definition aerial photos for now — I just want to see whether it can actually understand layout details.

After taking the job, Hy4 first broke down the image's color scheme, fonts, and layout one by one, even measuring how many pixels wide the left sidebar was, and only then started writing code to build the page and playback functionality together.

When the cropping tool hit a slight deviation midway, it switched approaches and directly analyzed pixels to locate lines and extract colors.

Alright, it even brings its own ruler and color swatches to the task.

The moment I saw the finished product, I stared at the screen for a full half minute. No wait, this has to be 99.9% faithful —

Beige base, orange-red accents, yellow-green color scheme, plus a thick bold headline and a vinyl record, and the retro vibe hits you immediately.

The navigation, playlist, playback area, and up-next list are all exactly where they should be.

I put the buyer's photo and the seller's photo side by side and searched for differences for ages, and the most noticeable thing turned out to be those two forward/back buttons in the top-left corner: one round, one square...

Honestly, without a magnifying glass, you wouldn't even know where to nitpick.

I looked through Hy4's delivery notes: play/pause, previous/next track switching, progress dragging, and volume adjustment are all wired up, and loop, shuffle, search, and favorite syncing are not left out either.

It even accounted for the case where you're too lazy to touch the mouse: spacebar controls playback, arrow keys adjust progress.

But when I clicked the play button, I discovered that it had even prepared three test tracks.

I didn't upload any audio files to it at all!

Video address: https://mp.weixin.qq.com/s/OThj_29CAuHn6ioaqjkSpQ

It turns out these tracks were synthesized in real time by it using the browser's native Web Audio, so I didn't have to rummage through my whole computer for audio files — just open the webpage and listen.

Good, good, good. I only gave it one image, and Hy4 not only delivered the player but also threw in a test track library, conveniently taking care of the sound check too.

Then I went through its self-test logs.

According to the process in the records, after writing the code, Hy4 quietly spun up a headless Chrome browser in the background and, in a sandbox environment, clicked and dragged each UI button one by one, taking screenshots along the way for visual diff comparison.

The first round didn't actually pass everything in one go:

48 of 51 items passed, two got stuck on the test script, and the last one was caught by Hy4 itself — after the list finished the last song, the logic for clicking next wasn't fully closed out.

It promptly fixed this edge case and reran the regression.

The final report showed all 59 tests passed, with zero console errors and zero runtime exceptions.

A 2D interface alone wasn't satisfying enough, so I threw Hy4 an amusement park floor plan and asked it to "stand" the park up.

Where the facilities are and how big they are, how the paths curve — all had to follow the image; the Ferris wheel and carousel couldn't just stand there posing, they had to spin; and this amusement park also had to let me walk around inside it.

After a series of geometry modeling, shader configuration, and camera rigging, it actually raised the entire "Starlight Amusement Park" from the ground up on the spot.

The classic red-and-white carousel canopy, the snack stands along the main path, and the low-poly trees and grass distributed along the streets were all projected into three-dimensional space to scale.

The low-poly art style is quite refreshing too—it looks like a small game perfect for wandering around in for a couple of laps while slacking off.

I had already prepared myself to get lost right after entering, but surprisingly there was a small minimap in the top-right corner: facilities, roads, lawns, and even where I was currently standing were all marked out.

After entering the park, WASD to move, drag the mouse to turn, hold Shift to speed up, and open labels when you want to look at facilities.

If you get dizzy from wandering, you can return to the entrance with one click—the whole point is being able to find your bearings at any time.

It even prepared an "auto tour along roads" button; click it once, and you don't even have to walk the paths yourself.

Well then—ask it to build an amusement park, and the guided tour service opens for business along with it.

Video address: https://mp.weixin.qq.com/s/OThj_29CAuHn6ioaqjkSpQ

Agent delivery capability maxed out

After touring the amusement park, let's get back to some real work and try out the Agent and long-horizon task capabilities that Hy4 specifically strengthened this time.

I fed it a product information package for the "Shiguang Beam Go 2" portable smart projector and asked it to make a new-product introduction PPT of about 15 pages.

Hy4 first went through the materials from beginning to end. It didn't randomly source images either—it directly used native code to draw vector diagrams with a highly consistent style, while charts were kept in editable form for easy later modification.

When rendering page 4, it found that a long block of text had overflowed the page's safe border. It immediately restructured the layout in the background, splitting the single column into a card flow, and only continued after a second validation passed.

In the end it delivered 16 pages of PPT. Fine, one extra page as a bonus.

But office workers all know that once the PPT is done is often exactly when the PPT revision begins.

I then threw it a dark, high-contrast template with a completely different visual system and asked it to redo everything according to the new layout.

Hy4 didn't rush to change all 16 pages. It first picked one page to try fitting into the template, confirmed there were no problems, and then batch-modified the remaining pages.

I have to say, the ability to work while validating is a very standout advantage this time. It doesn't just dump everything out and call it done; instead, it tries in small steps, advances only after validation passes, and greatly reduces rework.

Next, I gave Hy4 another job:

build a highly usable batch image processing tool.

Cropping, resizing, compressing, exporting—doing a single image a few times is fine, but going through a batch one by one, the mouse may not break, but the person does first.

This time I simply handed the repetitive manual operations over to it.

After receiving the task, Hy4 first studied the excellent designs of industry open-source benchmarks Squoosh and Image Workspace in the environment, drawing on technical approaches such as the original/compressed comparison slider and preset workflows.

Then it quickly broke things down into the upload flow, processing flow, Web Worker multithreaded parallelism, Canvas graphics pipeline, and ZIP archiving module, and started coding directly.

When the finished product was delivered, it even had a name—

ATELIER Image Batch Processing Workbench.

Off-white background, orange accents, and a three-step guide pops up when you open it. The slogan is: "Leave the repetitive to the tool, keep the aesthetic for yourself."

Fine, I'll handle the aesthetics, it handles the repetition.

According to the delivery design, images can be dragged in as a whole batch or selected directly by folder; on the left, you uniformly set cropping, dimensions, compression, and format, and the few images that need "special care" can be adjusted individually.

It also conveniently built in common presets such as 1 social avatars, e-commerce hero images, and extremely small file sizes. Dragging the slider lets you see quality loss in real time, and finally one click packages everything into a ZIP for browser download.

Video address: https://mp.weixin.qq.com/s/OThj_29CAuHn6ioaqjkSpQ

In addition, Hy4 can now really pick up the mouse and work.

Open JS Paint in Chrome, use the mouse to copy the Hunyuan kiss-kiss meme sticker, and record the operation process.

Video address?: https://mp.weixin.qq.com/s/OThj_29CAuHn6ioaqjkSpQ

Looking at the screen, finding tools, choosing colors, and then controlling the mouse to put pen to canvas—what it's using is precisely its newly supported CUA (Computer Use) capability.

Well then—other AIs are busy generating images with one click, while Hy4 is here practicing mouse drawing. Above, a reference image tests its eyesight; below, drawing software tests its craftsmanship (doge).

More power, same price

After playing around with all this in hands-on testing, my biggest takeaway is that Hy4's autonomous execution completion rate on long-horizon tasks has improved a lot.

When facing multi-step tasks, it can continuously maintain goals and task state, perform staged self-checks during execution, locate problems based on runtime results, modify the plan, and re-validate until it can deliver bug-free results.

The cases I described above were indeed all one-shot deliveries after its autonomous execution, checking, and modification.

This round of training for Hy4 also places greater emphasis on delivering value in real productivity scenarios.

Front-end optimization has been refined down to complex interactions, business logic, state management, and front-end/back-end integration;

Targeted training has also been conducted for professional software such as Blender and CAD, as well as long-horizon cross-application tasks, strengthening task planning and execution capabilities.

In office scenarios, Hy4 has also carried out targeted optimization of the search pipeline around time-sensitive retrieval, multi-source information integration, and factual consistency verification.

In internal testing, Hy4's hallucination rate dropped from 12.95% in the preview version to 7.2%, and its speculation-type error rate fell from 23.1% to 5.6%;

On the finance side, Hy4 also completed a dedicated round of professional coursework.

The training covered directions including fund research, securities analysis, quantitative strategies, asset allocation, corporate finance, and tax accounting, building thousands of workspaces and tens of thousands of high-quality tasks based on real scenarios, so the model could learn professional workflows and standards.

A total of 213 experts participated in the public beta testing, 70% of whom came from leading external financial institutions and have more than five years of work experience —

These graders are not so easy to fool.

As a result, across 579 real financial tasks, Hy4 ranked first among open-source models in overall performance.

It seems this round of "further study" really paid off.

The span of scientific research tasks is also expanding.

Hy4 combines complex reasoning, scientific programming, and long-horizon execution, demonstrating research capabilities in areas such as mathematical construction and physical modeling in the internal TenureTrail benchmark covering 7 major disciplines and 123 executable research tasks.

Faced with a dataset containing 6.99 million single cells and 84 two-dimensional unordered mouse embryo sections without any temporal or spatial continuity markers, mixed with noise data such as mirror flips commonly found in experimental sections; under severe conditions of offline operation, no dedicated registration tools, and no cell type annotations, Hy4 took 1.5 hours to autonomously complete spatial alignment and 3D mesh reconstruction of mouse embryo sections.

With zero manual intervention throughout the entire process, the final alignment precision even beat the default configuration workflow of professional registration tools.

It has also begun to string together complex reasoning, scientific programming, and long-horizon execution.

In research conducted in collaboration with the mathematics team at Peking University, Hy4 found a counterexample to the Méndez–Pinto–Villarreal conjecture, further proposed new conditions to refine the conjecture, and carried out Lean formal verification of the relevant proof.

These capabilities can now be experienced through WorkBuddy and CodeBuddy. From now until 23 on November 15, there is also a limited-time free trial!

Friends, if you still have any tasks that have been sitting around for a long time and always seem troublesome, you can arrange for Hy4 to handle them during this period~

One final reminder: Hy4 flash, which focuses on extremely low inference cost and is suitable for on-device and high-concurrency Agent orchestration, is also already on the way.