Tencent Hy4 vs DeepSeek, Kimi and Claude Opus 5.5: Tested

What is Tencent Hy4?

Hy4 is a large language reasoning model from Tencent's Hy team, which used to be called Hunyuan. It features 770 billion parameters in a mixture-of-experts architecture and a context window of 1 million tokens. It also supports vision, tool calling, and computer use. Currently, the weights of the final version aren't public, but Tencent plans to open-source them later, with support for vLLM and SGLang. Hy4's final release also hallucinates less than the preview, as it makes up facts in 5.45% of answers, while the preview did so in 15.76%.

Hy4 benchmarks

Here's how Hy4 compares against other models in the same class. The data below comes from Tencent.

Benchmark	Hy4	DeepSeek V4.1 Flash	Kimi K3.1 Pro	Claude Opus 5.5
DeepSWE v1.1 (coding)	74.9	73.8	76.3	72.6
SWE-Marathon v1.1 (long coding tasks)	43.8	46.3	53.8	44.4
Terminal-Bench 4.0 (terminal tasks)	44.4	40.0	66.8	66.4
APEX-Agents v1.1 (office agent tasks)	75.2	72.9	68.2	74.8
Agents' Last Exam v1.1	41.0	35.3	43.4	43.4
OSWorld 2.1 (computer use)	78.9	68.4	79.6	81.8
MMMU-Pro (image understanding)	78.6	76.7	83.4	88.2
ZeroBench (visual puzzles, pass@5)	55.4	51.0	56.9	63.5
Humanity's Last Exam (with tools)	59.1	57.1	61.8	66.2

To summarize that, Hy4 is ahead of DeepSeek V4.1 Flash on most tests and leads all four models on APEX-Agents, a benchmark of office tasks such as working with documents and spreadsheets. On the rest, Kimi K3.1 Pro and Claude Opus 5.5 mostly score higher. The gap is largest on Terminal-Bench 4.0, where both are more than 20 points ahead.

Hy4 vs DeepSeek V4.1 Flash vs Kimi K3.1 Pro vs Claude Opus 5.5: coding test

Now let's look at the tests. In the video, Hy4 is in the top left, DeepSeek V4.1 Flash in the top right, Kimi K3.1 Pro in the bottom left and Claude Opus 5.5 in the bottom right.

City growth timelapse

For the first task, we gave the models a challenge to create a 40-second timelapse in Three.js showing how a city grows in an empty valley — like a Cities: Skylines animation.

City growth timelapses written by Hy4, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Claude Opus 5.5

The prompt:

Create a single HTML file with a cinematic city-growth timelapse in Three.js. Load it with an import map from https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js (addons, if you need them, from https://cdn.jsdelivr.net/npm/three@0.160.0/examples/jsm/); no other libraries, no external models or textures: everything is procedural.

The timelapse plays as one 40-second story:
1. Dawn over an empty green valley with a river and hills.
2. Roads grow out from a crossroads as a branching network that follows the terrain, with a bridge over the river.
3. Buildings rise along the roads block by block and floor by floor: small houses first, then apartment blocks, and finally glass skyscrapers in the center. Each district has its own style, and parks and trees fill the gaps.
4. Cars drive along the roads and stop at traffic lights, and there are more of them as the city grows. A train runs across the city on an elevated track.
5. The sun crosses the sky, shadows sweep over the city and clouds drift. At dusk the windows and streetlights switch on one by one, headlights leave light trails, and the night has a starry sky.

The camera flies a slow cinematic path: it starts high above the valley and moves lower between the towers as they grow. Use InstancedMesh for buildings, trees and cars so it stays fast with thousands of objects. Add soft shadows, fog, and bloom on the lights at night. When the story ends, start again from a new random seed.

It must start automatically and run smoothly. Return only the complete code in a single ```html code block.

This test is about whether a model can plan a big procedural scene with thousands of objects, a day-night cycle and a moving camera in one file and keep it running smoothly.

Claude Opus 5.5 created the most detailed result that, in our opinion, looks like a big leap compared to all of the other models. Kimi K3.1 Pro also performed well. Unfortunately, the scene created by Hy4 looks the most blocky and would need several iterations of polish.

Weather app from a sketch

Since vision is new in Hy4's final release, we gave the model a task that tests how well it can interpret a design and replicate it in code. To do that, we gave the models a photo of a notebook page with a hand-drawn weather app screen. Around it are the designer's notes on colors, fonts and behavior, for example, that the °C / °F switch must change all temperatures and that rain should fall in the background when it rains.

Here's the input:

A notebook page with a hand-drawn weather app screen and design notes

And the prompt:

I sketched this weather app screen in my notebook. Build it as a single working HTML file (no external libraries) that matches the sketch and follows every note on the page. Use the weather data from the sketch. Return only the complete code in a single ```html code block.

And here's the result:

Weather apps built from the sketch by Hy4, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Claude Opus 5.5

Interestingly, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Claude Opus 5.5 built working apps with the data from the sketch: 18° and light rain in Lisbon, the hourly forecast from Now to 4 PM and seven days from Monday to Sunday. In all three, the °C / °F switch changes every temperature on the page, including the chart axis, and rain falls in the background. Kimi K3.1 Pro and Claude Opus 5.5 also put the temperatures on the chart points, as the sketch shows, but in both apps, the bottom menu covers the 7-day forecast on the first screen. DeepSeek V4.1 Flash fit the whole screen without overlaps in a minute and a half, but its chart has no labels on the points.

Hy4 didn't finish this task. In each of the three tries, it reasoned for 11 to 13 minutes. The second try got the furthest: its layout follows the sketch closely, and nothing covers the forecast. However, the answer ended in the middle of the script, so the chart stays empty and the °C / °F switch doesn't work. In the other two tries, the answer stopped in the styles or never got past reasoning.

Vision test

For the vision test, we made two images and worked out the correct answers in advance. To make the results more reliable, we ran every image three times.

Task, correct runs	Hy4	DeepSeek V4.1 Flash	Kimi K3.1 Pro	Claude Opus 5.5
Kayak on a roof	3 of 3	3 of 3	3 of 3	3 of 3
Station clock	3 of 3	3 of 3	3 of 3	3 of 3

Finding a kayak on a roof

In this test, the models need to find a yellow kayak in a drone photo of a dense old town, and count the swimming pools. Here's the input:

A drone photo of an old town with a yellow kayak on one roof

And the prompt:

This is a drone photo of a city taken from above.
1. Somewhere on a rooftop there is a yellow kayak. Where exactly is it in the photo, and what is right next to it?
2. How many swimming pools are on the rooftops, and where is each one?

All four models found the kayak on a large flat roof just right of the center and counted three pools in every run. In two runs, Hy4 also accurately gave the position of the kayak and every pool in percent of the frame: the kayak is about 54% across and 38% down. In terms of time, Claude Opus 5.5 and DeepSeek V4.1 Flash answered in about 10 seconds, while Hy4 took about three minutes.

Station clock and departure board

The last image is a station hall with an analog clock without numbers and a departure board. To answer, a model has to read the time from the hands, add six minutes of walking and check which trains to Lyon are late or cancelled. The clock shows 14:38, so the 14:41 train leaves too early, but you can still catch the 14:35 train, which is 15 minutes late.

Here's the input:

A station hall with an analog clock and a departure board

And the prompt:

I'm standing under this clock and need to get to Lyon. It takes me 6 minutes to walk to any platform. What time is it now, which train should I take and from which platform?

All four models chose the right train, TGV 6623 from platform 11, in every run, but not all were able to read the clock correctly. Kimi K3.1 Pro and Claude Opus 5.5 read 14:38 every time, and Hy4 did so in two runs and answered "about 14:38–14:40" in the third. DeepSeek V4.1 Flash read the clock as 14:40 in two runs out of three. Hy4 and Claude Opus 5.5 also mentioned the 14:52 OUIGO, but only Opus explained why it's a worse choice: it goes to Lyon Saint-Exupéry, the airport station outside the city.

How we tested

Every test result was generated from a single prompt. Every model used its default reasoning settings. In the coding tasks, each model had to return one HTML file that runs in Chrome without any setup. All tests except Claude Opus 5.5's timelapse were run in Overchat AI.

Hy4 reasoning speed compared

Hy4 isn't the fastest model. Claude Opus 5.5 and DeepSeek V4.1 Flash answered the vision tests the fastest, and DeepSeek was also the fastest in every coding task. On average, Hy4 spent two and a half to a little over three minutes on each image, 13 to 18 times longer than Opus. Kimi K3.1 Pro was the slowest in coding: it spent 16 to 20 minutes on each program.

Task	Hy4	DeepSeek V4.1 Flash	Kimi K3.1 Pro	Claude Opus 5.5
City growth timelapse	4 min 18 s	2 min 46 s	19 min 39 s	—
Weather app from a sketch	11–13 min, unfinished	1 min 36 s	15 min 53 s	5 min 40 s
Kayak on a roof, average of 3 runs	3 min 14 s	9 s	2 min 31 s	11 s
Station clock, average of 3 runs	2 min 37 s	23 s	53 s	12 s

Tencent Hy4 API price

Hy4 costs $0.834 per million input tokens and $2.50 per million output tokens.

Model	Input, per 1M tokens	Output, per 1M tokens
Hy4	$0.834	$2.50
DeepSeek V4.1 Flash	$0.15	$0.60
Kimi K3.1 Pro	$2	$10
Claude Opus 5.5	$4	$20

In Overchat AI, a message to Hy4 costs 2 credits. For comparison, a message to DeepSeek V4.1 Flash costs 1 credit, Kimi K3.1 Pro 16 credits and Claude Opus 5.5 110 credits.

Hy4 vs Claude Opus 5.5: which one to use?

In the vision tests, Hy4 was as accurate as Claude Opus 5.5, but at the same time, it was quite slow, taking 13 to 18 times longer to answer than Opus. That's a negative, but it's offset by a much lower price. In plain words, when you need a cost-efficient model to analyze an image or a series of images, or work with visual media but you're not necessarily concerned with speed, Hy4 becomes one of the best available choices.

The story is different for coding, where Hy4 is not as strong as frontier models. In our test, for example, its scene had more bugs than the ones from Kimi K3.1 Pro and Claude Opus 5.5, and while preparing for the tests, we had multiple occasions when it would get stuck in reasoning loops.

How to use Hy4 in Overchat AI

You can use Hy4 in Overchat AI on a paid plan:

In the message box, switch to Chat mode and click the name of the current model.
Choose Hy4 in the Model list.
Type your prompt or attach an image, then press the send button.

FAQ

These are the questions people ask most often about Hy4.

What is Hy4?

Hy4 is a large language model from Tencent's Hy team, formerly called Hunyuan. It's a reasoning model with 770 billion parameters and a context window of 1 million tokens. It accepts text and images, answers with text and can operate a computer.

When was Hy4 released?

Tencent released Hy4 on November 4, 2026. Hy4 preview, a text-only version of the model, came out at the end of August 2026.

Is Hy4 free?

Hy4 preview has open weights under the Apache 2.0 license, so you can download it for free. The weights of the final version aren't public yet, as Tencent plans to release them later. Through the API, Hy4 costs $0.834 per million input tokens and $2.50 per million output tokens.

Is Hy4 better than DeepSeek V4.1 Flash?

On Tencent's benchmarks, Hy4 is ahead of DeepSeek V4.1 Flash on most tests. In this comparison, both models solved every vision task, and DeepSeek answered 7 to 22 times faster, although it misread the station clock in two runs. In coding, each of them wrote one working program out of two.

Is Hy4 better than Claude Opus 5.5?

No. On Tencent's own benchmark table, Claude Opus 5.5 scores higher on most tests. In this comparison, Hy4 matched Opus in the vision tests but took much longer to answer, and it wrote fewer working programs.

Where can I use Hy4?

Hy4 is available through the Tencent Cloud API, OpenRouter and Vercel AI Gateway, in Tencent's apps such as Yuanbao, CodeBuddy and ima, and in third-party apps such as Overchat AI.