Article title: Hands-on with Tencent Hy4 flash: For two mao, AI can build a 3D interactive webpage from scratch
Body:

Hands-on with Tencent Hy4 flash: For two mao, AI can build a 3D interactive webpage from scratch
2026-11-16 17:53
In early November, Tencent Hy4 was released, and the new model jumped to first place among open-source models on the large model arena.
The three models ahead of Hy4 are GPT-6 Sol (Max), Claude Opus 5 (Max), and Fable 5 (High), with a group of strong closed-source models nearby such as GPT-6.1 Terra and GPT-5.6 Sol.

We previously experienced Hy4's long-horizon Agent tasks and everyday office capabilities in WorkBuddy. From frame-by-frame video analysis, recreating games and webpages, to file organization and Office document processing, Hy4 delivered solid results.
The overall hands-on performance also basically matches its results on the large model arena.

But for users who frequently use WorkBuddy, model capability is only one of the selection criteria. Whether it is free, whether queuing is required, and whether execution speed is fast enough also affect the actual experience.
For example, when opening WorkBuddy during weekday peak hours, DeepSeek V4.1 Flash, with a multiplier of only 0.11x, often still has a persistently high busy level.

As early as the release of Hy4, Tencent had previewed an efficiency-focused Hy4 flash.
Just now, Hy4 flash was officially released. Tencent hopes to bring Hy4's existing Agent, long-horizon task, and productivity capabilities to high-frequency usage scenarios at a lower cost.
APPSO received early access to test Hy4 flash. What we cared about most this time is whether it can reduce time and Token consumption during Agent execution while ensuring task completion quality.

Compared with Hy4, the new Hy4 flash has undergone extensive optimization for Token efficiency. Hy4 flash has 224B total parameters, 12B activated parameters, and supports multimodality and a 1M context length. Model output is more concise, and throughput and overall response speed have also improved.
Even more surprising is that on WorkIndex, Hy4 flash's single-task cost is only 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (max tier, busy-hour price).


Higher Token efficiency also makes Hy4 flash considerably cheaper than Hy4. Hy4 flash's input price is 0.5 yuan per million Tokens input, while output is 2 yuan per million Tokens output, almost one-tenth of Hy4's price.

We can directly select the Hy4 flash model in WorkBuddy. The new model also comes with a limited-time free benefit for two weeks, from November 16 to November 30, with free quota granted daily.
Faster and cheaper than DeepSeek
Today's Agents often take at least half an hour to get work done.
Open the execution record and watch it read files, break down tasks, write code, call tools, and check results... After a full workflow, tens of minutes have passed, and the model may still be repeatedly checking the same file, or even reintroducing a problem it just fixed.
These detours do not appear in the final result screenshots, but they do materially increase task time and the Token bill.

So when Agents start taking on real work, we also need to care about a more practical question: how much time and cost does completing the same thing actually require?
To test the actual efficiency of Hy4 flash, we prepared the same tasks in WorkBuddy and assigned them to Hy4 flash and DeepSeek V4.1 Flash respectively. During testing, the prompts, tool permissions, and other settings remained consistent, with a focus on observing the two models' execution process, final delivery, and the time and Token consumption behind them.
We first used some repetitive, trivial tasks within WorkBuddy that nevertheless take time. Although these tasks do not require the model to develop a complex system from scratch, they still test its ability to understand and organize information and call tools.
For example, conducting some deep research tasks to see whether it can efficiently obtain results that are reliable, clearly organized, and easy to use further.

We asked it to conduct deep research on the recent matter of 黄仁勋 resigning as CEO of 英伟达 due to an affair, and to present me with a webpage analysis report. The content delivered by DeepSeek V4.1 Flash was overall similar, also presenting the course of the matter from several aspects such as summary, timeline, duration, fundamentals, products, and Wall Street, but it took 10 more minutes than Hy4 flash.

The Token consumption of the two research reports also shows that DeepSeek V4.1 Flash used nearly 180,000 Tokens, with a total cost of about 0.56 yuan.

Looking at programming next, we asked the two models to build an interactive 3D historical sailing ship in WorkBuddy.
Hy4 flash quickly built the basic structure of the sailing ship, and the webpage could interact normally. Overall it was a relatively complete prototype.

DeepSeek V4.1 Flash did it more meticulously. The water surface fluctuations were more natural, and the sailing ship also had more details and content to explore.

However, the gap in completion time between the two models was large. DeepSeek took 1 hour 30 minutes, while Hy4 flash took only half an hour.
For users who just want to quickly validate ideas and build webpage prototypes, especially in scenarios requiring faster iteration, Hy4 flash is more suitable for high-frequency use, and its output already has a foundation for further modification.
DeepSeek V4.1 Flash consumed about 300,000 Tokens in this round of tasks, plus about 16 million Tokens of cache hits, with an API cost of about 1.21 yuan.

Hy4 flash's conversation data recorded about 149,000 Tokens consumed and about 4 million Tokens of cache hits. According to Hy4 flash's API pricing, completing this interactive 3D historical sailing ship is estimated to cost 0.23 yuan.

In addition to the above cases, we also counted the performance of the two models in 10 coding and office tasks. Judging from the final results, the two sides were overall at a similar level. But in terms of first-round delivery time, Hy4 flash's advantage was more obvious.

Values written like {{NAME_ab12cd34}} are placeholders for real values. Use them exactly as written, in replies and in tool calls alike: they are filled in with the real values before anything runs. Don't ask to verify them, and don't replace them with examples.
After this round of testing, we believe that Hy4 flash and DeepSeek V4.1 Flash are roughly at the same level in terms of task completion capability. DeepSeek sometimes invests more tokens, and in some scenarios it does indeed produce richer details.
Hy4 flash's advantages are also obvious: in the process of completing actual work, the model can spend less overhead and provide faster responses. Especially when some work requirements are clear and the format is fixed, as long as the model accurately understands and smoothly executes, lower waiting time is very meaningful in these cases.
For high-frequency productivity tasks such as deep search, daily office work, and file organization, we would recommend Hy4 flash. It can deliver decent-quality results at a lower cost in both time and tokens. Even for creative tasks such as web prototypes and mini-game development, you can first use Hy4 flash to quickly generate a first version, validate the idea, and then decide how to improve it afterward.

For high-difficulty tasks that require sustained reasoning and repeated verification, such as long-horizon software engineering and modifications to complex codebases, we still recommend the previously released Hy4. Investing a bit more time and tokens in exchange for stronger complex-task handling capability is often worth it.
This also gives Hy4 and Hy4 flash a clearer division of labor. Daily high-frequency tasks can be confidently handed to Hy4 flash, and when more complex work comes up, switch to Hy4. We can choose among task quality, execution speed, and usage cost based on actual needs.
A more practical coding and office companion
The upper limit of capability determines what work a model can take on, while execution efficiency affects how many times these capabilities can be used. This time we continued to test Hy4 flash in coding and office scenarios to see what it can achieve.
For example, on the office side, we tested PPT creation, file organization, and conversion between multiple document formats.

Inside WorkBuddy, we sent it a paper and asked it to convert this pdf into a markdown document, a word document, and a PPT for a group meeting presentation, with all content provided in both Chinese and English, and to translate it for us.


Another scenario worth trying is having Hy4 flash collaborate with WorkBuddy's expert team.

Normally, when using the expert team inside WorkBuddy, the system usually reminds us that multi-Agent collaboration may consume more credits. After all, a task needs to be completed by different Agents working in division of labor, and the number of model calls involved behind the scenes will increase accordingly.
Hy4 flash's low multiplier makes the cost of using such complex workflows more acceptable. We used the deep research team to help investigate what fable 5.2's capabilities are actually like now and what it can do.

For some coding tasks, by referencing a one-sentence prompt from netizens on social media, we were able to make a "spot the difference" mini-game.

After that, we increased the difficulty and asked it to develop from scratch, using three.js, a single-player 3D top-down fantasy RPG game, and to continuously improve the gameplay and content across multiple rounds of dialogue.
In this simple game, we can explore a large map of about 400×400: the village, Broken Ridge Mountain, Sunken Fruit Orchard, Whispering Swamp, Sunken Ancient Road, and the final "Land of Blankness."
Like common RPG games, players can talk to NPCs, light up survey markers and large anchor points, and collect 6 memory fragments. There is even combat against the "Unformed" using melee, dodging, and four types of magic unlocked progressively, along with various storylines waiting to be unlocked.

Hy4 flash also inherits Hy4's visual understanding capability. During development, it can check pages against screenshots of generated results, and it can also understand requirements based on images and videos provided by users to continue adjusting the work.
From mini-games and 3D web pages to office documents and multi-Agent collaboration, Hy4 flash is already able to cover many real work needs.
For Hunyuan, this is also the meaning of launching a Flash model: to allow existing visual, reasoning, and Agent capabilities to be called repeatedly at lower cost.
The greatest value of Flash is letting everyone use AI with confidence
This is the first Flash model in Tencent's Hy series, but among domestic models, various Flash models focused on "kill lines" and "more, faster, better, and cheaper" are already common.
Model vendors are all competing on price and speed, eager to push the price per million tokens down to a few dimes. But when these models enter an Agent like WorkBuddy, we also find that looking only at the API price table, it is already hard to judge whether a model is actually cost-effective.
A cheap model that takes an hour and a half to complete the work is not necessarily more cost-effective than a model that can deliver in half an hour. Of course, if in pursuit of speed you still end up having to spend a lot of human time reworking, that is equally not worth it.

This is also a fairly obvious impression we had after experiencing Hy4 flash.
Its generation results are not always better than DeepSeek V4.1 Flash, and in some scenarios it has fewer details. But when facing a large number of daily office, web prototype, and Agent collaboration tasks, Hy4 flash is already able to deliver, in less time and at a lower price, results of comparable quality that can continue to be revised.
Moreover, when an Agent does work, it is often hard to get it perfect on the first try. The faster we get the first version of the result, the earlier we can find problems, adjust direction, and then let it continue completing the subsequent work.

Models are getting cheaper, and ultimately the benefit should be that we can use the same time and budget to let AI do a few more genuinely useful things.
We can confidently hand various big and small matters to AI and see what it can achieve, without first having to consider whether this task is really worth having AI do it.
When models, tools, and product scenarios like WorkBuddy enter the same iteration chain, the gap between Agents also becomes about who can let users wait a little less, watch a little less, and rework a little less, quietly finishing more real work.