AI News

Tencent Releases Hunyuan 4 flash: 224B Parameters, One-Million-Token Context, Hands-On Test Creating an MV and 3D Scenes

Phoenix TechnologyUpdated Nov 16 21:12

AI summary

Tencent released Hunyuan 4 flash (Hy4 flash), with 224B total parameters and 12B activated, supporting multimodality and a one-million-token context, a decoding speed exceeding 150 tokens/s, and API pricing of 0.5 yuan per million input tokens and 2 yuan per million output tokens. Tencent's blind test shows its user satisfaction score of 2.79 points, close to Claude Haiku 5.5's 2.81 points, with a per-task cost of 0.26 yuan, lower than DeepSeek-V4.1-Flash-1017's 1.20 yuan. The author's hands-on tests found that the model could successively complete a pixel-art MV of Jensen Huang, a simplified version of The Sims game, a 3D tennis scene, and a Blender animation of a pelican riding a bicycle, but image fineness, restoration of complex structures, and subtitle synchronization still show clear deficiencies.

Why read this

AI score 77

The author conducted multiple rounds of hands-on testing with an MV, a game, 3D scenes, and Blender animation, revealing the lightweight model's long-horizon task capabilities and its shortcomings in fine detail.

Topics

Tags

Body · AI translation

Exclusive hands-on test of Tencent Hy4 flash: it produced a Jensen Huang MV in one go, cheaper than DeepSeek. Zhi Dongxi download the client | Exclusive first look. November 16, 2026, 16:53:33, from Beijing

Jensen Huang, wearing his black leather jacket, was turned into a pixel character and took the stage in Jolin Tsai's "Sun Never Sets." When it came to the line "postcard of love," Jensen Huang pulled out a green GPU chip.

The familiar black leather jacket and the familiar moves do capture a bit of the Nvidia GTC flavor.

This MV is of course hardly polished. Jensen Huang's movements are somewhat stiff, the visual details are rather rough, and even the lyric subtitles fail to line up accurately with the background music.

But the entire video was made by AI through writing code, and even the pixel character Jensen Huang was drawn by the AI itself. The model behind it is Tencent Hunyuan 4 flash, just unveiled by Tencent's Yao Shunyu team.

Zhi Dongxi news, November 16: just now, Tencent officially released its new-generation model Hunyuan 4 flash (Hy4 flash), with 224B total parameters (224 billion), 12B activated parameters (12 billion), support for multimodality and a 1-million-token context, and a decoding speed exceeding 150 tokens/s.

This model focuses on low cost. Its API price is only 0.5 yuan per million input tokens and 2 yuan for output, which are 1/12 and 1/9 of the official Hunyuan 4 version respectively, and the cached-hit input price is as low as 0.02 yuan.

In Tencent's published evaluation of everyday work tasks, Hunyuan 4 flash outperforms the Hunyuan 4 preview version, with an average cost of about 0.2 yuan per task.

Another blind test involving 85 internal users and 259 work tasks showed that Hunyuan 4 flash's user satisfaction is close to Claude Haiku 5.5, with an average cost of only 0.26 yuan per task, lower than DeepSeek-V4.1-Flash-1017's 1.2 yuan, and with shorter task times.

To test its actual capabilities, we conducted a round of hands-on testing in WorkBuddy first. Besides the Jensen Huang MV at the beginning, we also had it develop a "The Sims 5" game, build a 3D tennis match, create an animation of Beijing landmarks, and call Blender to make a pelican ride a bicycle.

1. Having Jensen Huang shoot an MV, completing more than 50 storyboard shots in one go

First, let's look at the generated Jensen Huang "Sun Never Sets" MV. This time, I referred to a tutorial on making music MVs with AI and had the model complete the steps in sequence: music analysis, art style design, character drawing, script storyboarding, and video production.

Among these, the character design also used the model's visual understanding capability. I uploaded a real photo of Jensen Huang and had the model refer to features such as the hairstyle, glasses, and black leather jacket in the photo to transform him into a chibi pixel character.

The model not only generated the character image but also specifically made an interactive character preview page. On this page, I can switch among 5 orientations, 12 expressions, and 12 actions for Jensen Huang, and also have him pick up a GPU chip, a microphone, or a laptop.

As can be seen, the character's gray-black short hair, glasses, and black leather jacket are all retained, and the lapels and silver-gray zipper on the clothes are also visible. The overall image is fairly recognizable.

Next, the model designed the plot according to the rhythm and lyrics of "Sun Never Sets" and generated an overview storyboard of more than fifty shots.

Finally, the model wrote animation code based on these storyboards and synthesized the MV at the beginning.

However, the final video still has several obvious shortcomings.

First, the visual fineness is insufficient, the character movements are somewhat stiff, and some shot transitions are not smooth enough.

The more obvious problem is subtitle misalignment. Although we explicitly asked the model to create subtitles according to the lyric timeline and to check synchronization during export, the final subtitles still failed to keep accurate pace with the music.

Overall, the model completed a whole set of long-horizon tasks from character design and storyboard planning to animation production and video export, but there is still room for further refinement in visual details and audio-visual synchronization.

2. Building "The Sims 5" and a 3D tennis court from a single sentence, and even using Blender directly

Besides the MV, I also tested Hunyuan 4 flash's ability to complete complex front-end development, 3D scene creation, and cross-tool tasks.

(1) Creating "The Sims 5" from a single sentence

I first gave the model a game production request: make a "The Sims 5" with graphics and gameplay.

Unexpectedly, the model actually produced a playable simplified version.

In the center of the screen is a small house in isometric view, with three family members inside. On the left, the characters' hunger, energy, fun, social, and hygiene status are displayed in real time, while on the right, character information can be viewed.

Players can click on furniture to arrange actions and can also turn on automatic living so that the characters act on their own. The top of the page provides options such as pause, speed-up, and build mode, and the characters' behavior is also recorded below.

Of course, this is far from the real "The Sims." The characters are just simple cartoon figures, the furniture is made up of rather basic geometric shapes, and the room structure and game graphics are very crude.

However, the model has already implemented basic gameplay such as changes in character needs, automatic behavior, and furniture interaction, rather than just making a static page that looks like a game.

(2) A 3D tennis match was built, but the players cannot hit the ball

Next, I gave the model a more difficult task: creating a complete 3D tennis match. The prompt required it not only to build the court, stands, club and other facilities, but also to implement player serves, rallies, match scoring, and switching between different court materials and camera angles.

In the end, the model produced an interactive 3D tennis game.

We can switch between panoramic, tracking, close-up and other camera angles, and also adjust the match speed, court type, and day/night scenes.

The entire court uses a pixel style, with stands and club facilities built around it, and the players and tennis ball in the match can also keep moving.

But on closer inspection, quite a few bugs can still be found. For example, flagpoles appear near the court in unreasonable positions; the two players seem to be rallying back and forth, but sometimes their rackets do not touch the tennis ball at all; and the bodies of some spectators in the stands are merged into their seats.

(3) Calling a Skill on GitHub, but Beijing landmarks came out distorted

We also tried having the model call the perspective illusion Skill on GitHub to create a 10-second animation themed on Beijing landmarks.

The model needed to place buildings such as the Hall of Prayer for Good Harvests at the Temple of Heaven, the Corner Tower of the Forbidden City, the CCTV Headquarters, China Zun, and the Bird's Nest at different depths of field, and have the camera pass through these landmarks in sequence.

The model successfully completed the video and even automatically added BGM with a flavor of traditional Chinese music.

But the building reconstruction was not ideal. The Hall of Prayer for Good Harvests at the Temple of Heaven and China Zun were basically recognizable, but the shape of the CCTV "Big Pants" was severely distorted, and some other landmarks were also hard to recognize as their original forms.

Although the camera movement and scene transitions basically met the requirements, the geometric structure of complex buildings remains a difficult point for the model.

(4) The rainy-night convenience store at the foot of Mount Fuji was made, and the atmosphere is quite good

Next, we asked the model to create a Japanese-style rainy-night convenience store street corner, presented overall as a 3D miniature landscape that can be freely rotated, dragged, and zoomed.

The visual effect this time was relatively good. Elements such as the convenience store, vending machines, bicycles, streetlights, and street-corner guardrails were basically all present. The pink-purple neon lights, rain streaks, and reflections on the wet road surface also created the rainy-night atmosphere of Japanese animation.

The model implemented free rotation and zooming, and we can view the entire miniature street scene from different angles. Compared with currently impressive AI 3D scene generation works, it still has a gap in artistic detail and overall aesthetics.

(5) Directly calling Blender, a pelican rode a bicycle

Finally, we tried having the model operate the professional 3D software Blender. This time, the task was to have it write a script through the Blender Python API to create an approximately 16-second animation of a pelican riding a bicycle.

The model successfully built the characters and scene, including the pelican, bicycle, and walkway.

In the video, the model also set up tracking, close-up and other shots as required.

However, the overall picture was rather rough. The pelican and bicycle were both made of simple geometric bodies, and some rendered frames had almost no material color, looking like untextured white models without artistic processing.

This was also related to our prompt: we specified the character actions and camera movement, but did not explicitly require scene color schemes and material styles.

The pelican's action of "pedaling" the bicycle was done well. But the pelican's wings did not touch the handlebars at all, while its two webbed feet could still move alternately with the pedals.

Throughout the task, the model spent about 26 minutes writing the Blender code, and then more than 20 minutes on subsequent video rendering, for a total of nearly 50 minutes. But we used a Mac, and the rendering time was also affected by the local computer configuration.

In the end, the model not only generated the Blender script, but also ran it in the software and exported the video. This at least shows that it can already control professional 3D software through code to complete basic animation production.

In this round of experience, the more notable strength of Hunyuan 4 flash was its ability to complete multiple steps in succession and connect different tools together.

But its shortcomings were also fairly clear: first, precision—the visuals, models, and materials mostly remain at the level of "recognizable"; second, aesthetics—the camera, color scheme, and composition largely rely on the prompt as a fallback; third, reconstruction of complex structures such as buildings, human bodies, and machinery—the more complex they are, the more likely they are to be distorted; fourth, the self-checking loop—the model still cannot find problems on its own.

As a model positioned as Flash, it can already complete many complex coding and creative tasks. If the generated content is to be used directly as a mature work, there are still many places that require manual debugging.

III. Cost per single task is less than 0.3 yuan! Blind test satisfaction close to Claude Haiku 5.5

In addition to the results we actually experienced, Tencent also published evaluation comparisons between Hunyuan 4 flash and models such as DeepSeek, Claude, and GLM, focusing on how much money is needed to complete the same task.

First, look at WorkIndex, Tencent's self-built evaluation set for daily work. This evaluation covers 8 categories of daily work tasks, including PPT creation and modification, document editing, spreadsheet repair, audio and video processing, and visual understanding.

On WorkIndex, Hunyuan 4 flash scored about 42, higher than Hunyuan 4 Preview's 39, but still lower than Hunyuan 4 Official's 51.

▲Performance and cost on internal and external tasks.

In terms of price, the average cost per task for Hunyuan 4 flash is about 0.2 yuan, about 3.2 yuan for Hunyuan 4 Preview, and about 5.7 yuan for Hunyuan 4 Official.

Breaking it down, on PPTGenBench, the PPT generation evaluation, Hunyuan 4 flash scored 54.2, higher than Hunyuan 4 Preview's 52.3; on PPTCoEditBench, the PPT collaborative editing evaluation, the two scored 74.0 and 71.5, respectively.

But when it comes to more complex software engineering tasks, there is still a large gap between flash and the official version. For example, on Terminal-Bench 4.0, Hunyuan 4 flash scored 19.2, while Hunyuan 4 Official scored 44.4; on ProgramBench, the two scored 16.0 and 30.0, respectively.

▲Detailed comparison of public Agent benchmarks and multimodal evaluations

Beyond benchmark scores, Tencent also organized a round of blind tests on real work tasks. This test invited 85 internal WorkBuddy users to evaluate 27 categories and a total of 259 daily work tasks, comparing user satisfaction with the model outputs, end-to-end task time, and cost.

In the user satisfaction evaluation with a maximum score of 4, Hunyuan 4 flash received 2.79, almost catching up to Claude Haiku 5.5's 2.81 and higher than GLM-5.3-Flash's 2.53, but lower than DeepSeek-V4.1-Flash-1017's 3.09.

The median end-to-end task duration for Hunyuan 4 flash was 6.7 minutes, the shortest among the four models. Claude Haiku 5.5 was 6.8 minutes, GLM-5.3-Flash was 7.6 minutes, and DeepSeek-V4.1-Flash-1017 was 10.3 minutes.

In terms of cost per task, Hunyuan 4 flash required only 0.26 yuan, Claude Haiku 5.5 was 4.73 yuan, DeepSeek-V4.1-Flash-1017 was 1.20 yuan, and GLM-5.3-Flash was 0.81 yuan.

In other words, in this round of blind testing, Hunyuan 4 flash's satisfaction differed from Claude Haiku 5.5 by only 0.02 points, yet its cost per task was only about 1/18 of the latter's.

IV. Input price only 1/12 of Hunyuan 4's: how does a 224-billion-parameter model cut costs?

The standard input price for Hunyuan 4 flash is 0.5 yuan per million tokens, with an output price of 2 yuan, and the input price for cache hits is only 0.02 yuan.

By comparison, the standard input price for Hunyuan 4 Official is 6 yuan per million tokens, with output at 18 yuan, and the input price for cache hits at 0.3 yuan. The three prices were reduced to 1/12, 1/9, and 1/15 of the original, respectively.

Tencent disclosed that Hunyuan 4 flash has a total parameter count of 224B, including 64B Engram parameters, but only activates 12B parameters per inference.

In the pre-training stage, Tencent introduced architectural designs such as hybrid attention, LatentMoE, and Engram, compressing the model's KV Cache size to 1/6 of Hunyuan 4 Official's and 1/24 of Hunyuan 3's, in order to reduce the memory overhead of long-context inference.

In the post-training stage, Tencent first improved model capabilities through Hunyuan 4 self-distillation, then used reinforcement learning to reduce the output length of end-to-end tasks by 50%, cutting down the number of tokens needed to generate to complete tasks.

For deployment, optimization methods such as Megakernel, 4-bit quantization, and dflash were adopted. Tencent says Hunyuan 4 flash's decoding speed has exceeded 150 Token/s, supporting large-scale low-cost deployment. Tencent plans to open-source Hunyuan 4 Flash and continue iterating on the model.

In terms of use cases, Tencent recommends the flash version for daily tasks such as deep search, document processing, PPT generation, simple front-end displays, and Computer Use; for more complex tasks such as long-horizon software engineering, it recommends the more capable Hunyuan 4 Official.

Conclusion: In the price war for lightweight models, Tencent does not want to be absent

If Hunyuan 4 Official is Tencent's "calling card" for benchmarking against top open-source flagships, then Flash is meant to direct effort toward cheaper, higher-frequency scenarios.

On this track, competitors are already lined up. DeepSeek V4.1 Flash is already an important player in the lightweight market.

In the blind test published by Tencent, DeepSeek-V4.1-Flash-1017's user satisfaction was still higher than Hunyuan 4 flash's, but the latter completed the test tasks with lower cost and shorter task time.

Combined with our hands-on testing, Hunyuan 4 flash can already complete some complex creative and development tasks through code, but the fineness, aesthetics, and self-checking ability of the generated content still fall short.

As lightweight models increasingly enter office, development, and content creation scenarios, whether a model can complete tasks at a sufficiently low cost is becoming another key point of competition among vendors.

"Special statement: The content above (including videos, pictures, and audio if any) is uploaded and posted by users of "Dafeng Hao," a self-media platform under Phoenix News. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services.”

Source

Phoenix Technology