Tencent Hy3 Real-World Test - Compared to GLM 5.1 and DeepSeek V4 Pro

Tencent recently released the Hy3, and WorkBuddy can directly...freeWhen it's in use, everyone's scrambling to use it, and there's a long queue.

Hy3 is an upgraded version of the Hy3 preview released at the end of April.

In real-world blind testing of the model, Hy3 outperformed GLM 5.1 on average. In inference,intelligentbodyIn tasks involving long contexts, its performance rivals that of larger flagship models both domestically and internationally (whose parameter scale is often 2 to 5 times that of Hy3).

Since the Hy3 is advertised as being comparable to larger (2-5 times) flagship models and outperforming the GLM 5.1 in actual tests, we should definitely test it based on this claim.

I selected two with higher specifications than the Hy3: GLM 5.1 and DeepSeek-V4-Pro, and one with similar specifications: DeepSeek-V4-Flash, for a side-by-side comparison.

Let's put it to the test and see what it's really like!

Case 1 Front-end Design

Prompt words:

Make a AI Model comparison Dashboard.

The content includes:

• 1. The top displays 4 models: hy3, GLM 5.1, DeepSeek V4 Pro, and DeepSeek V4 Flash.
• 2. The middle section is the scoring matrix: code, reasoning, writing, visual, speed, and stability.
• 3. The right side displays the advantages and disadvantages of the currently selected model.
• 4. The test case execution log is displayed at the bottom.

Design requirements:

• 1. Do not create a marketing landing page.
• 2. It should be like a professional SaaS tool, with high information density but not cluttered.
• 3. Supports desktop and mobile devices.
• 4. Real-world interaction: Switching models, filtering test types, and sorting scores.
• Avoid using large gradient backgrounds, glass-like designs, and hollow decorations.

Hy3:

Hy3 is highly complete, and its information architecture is most like a real comparative evaluation tool: top KPI, 4 model cards, scoring matrix, right-side model details, radar chart, dimension details, advantages and disadvantages, specifications area, test records and filtering are all included.

There is a noticeable blank area on the left side of the middle of the desktop, which causes an uneven browsing rhythm.

DeepSeek-V4-Pro:

DeepSeek-V4-Pro has a clear structure, with a 6-dimensional × 4 model scoring matrix. Selecting columns, filtering buttons, and test history tables all work.

However, the design lacks distinctiveness, resembling a "standard white-background SaaS table demo"; the matrix area has large blank spaces, and the page density is poorly controlled; in terms of details, it lacks model profiles, specifications, and graphical capability panels, resulting in insufficient memorability.

DeepSeek-V4-Flash:

DeepSeek-V4-Flash's cards, bar charts, and tables are all very stable, with a restrained visual style and no unnecessary decorations.

However, the so-called "scoring matrix" only displays six dimensions of the currently selected model, not a true four-model horizontal matrix. It has limited information hierarchy, resembling a model details page plus a record table, rather than a complete comparative evaluation workbench.

GLM 5.1:

The GLM 5.1 page is not running. The browser reports an "Invalid or unexpected token" error. The model cards, matrix, details, and test records are all not rendered, leaving only a dark shell and an empty header.

Ranking of this case: Hy3 > DeepSeek-V4-Pro > DeepSeek-V4-Flash > GLM 5.1

Case 2 3D Physical Interaction

Prompt words:

Create a 3D mini fan palace.

Require:

1. Use Three.js.
2. The image shows a tilted maze board and a small ball.
3. Use the arrow keys on your keyboard to control the tilt of the maze board.
4. The ball will roll depending on the direction of tilt.
5. There are walls, a finish line, and a failure hole.
6. The checkpoint time is displayed upon reaching the destination.
7. Reset after falling into the hole.
8. It needs to have basic physical effects; it can't just be a matter of changing coordinates to fake scrolling.

Very stable, comes with three.min.js pre-installed and can run directly; the physics implementation is clear, including tilting boards, rolling balls, wall collisions, falling through holes and resetting, finish line, timer, number of falls, and a D-pad for mobile devices. The ball's position and timer change noticeably after pressing a button.

DeepSeek-V4-Pro:

DeepSeek-V4-Pro has a very advanced approach, using React Three Fiber + Rapier, and the build process also succeeded, indicating that a physics engine should be used.

However, it has some major problems when actually playing. The left and right gravity logic is chaotic, and the board tilts up but then moves upwards.

DeepSeek-V4-Flash:

It features a canvas, a maze, a timer, openings, and a pop-up window for completing the level.

The downsides are that the game system is still somewhat weak: there are few openings, limited HUD information, and the left and right gravity logic is chaotic; the board tilts up but then moves upwards.

GLM 5.1:

The game is very well-made, featuring a complete maze, walls, multiple failed exits, a finish line, shadows, a HUD, a best time indicator, WASD/arrow keys, R to restart, and mouse view rotation.

However, the gravity logic in all directions is chaotic; the board tilts up but moves upwards.

In this case, the ranking is: Hy3 > GLM 5.1 > DeepSeek-V4-Pro > DeepSeek-V4-Flash

Case 3: Tower Defense Games

Prompt words:

Make a mini tower defense game.

Require:

1. There are fixed paths on the map, and enemies move along those paths.
2. Players can place turrets in open areas.
3. turretautomaticAttack the nearest enemy within range.
4. Enemies drop gold coins when they die.
5. The number and speed of each wave of enemies increase.
6. It includes health points, coins, waves, pause, and restart.
7. There are at least two types of turrets: high single-target damage and low area-of-effect damage.
8. The game needs to be balanced so that at least 5 waves can be played.

Hy3:

Hy3 features three types of towers: arrow towers, cannon towers, and ice towers, covering single-target, area-of-effect, and slowing effects; it also includes features like sell-for-refund, pause, restart, and wave status. The UI is clear, with good feedback on enemy health bars, paths, and tower positions, making it feel like a complete mini-game prototype.

DeepSeek-V4-Pro:

The React/Vite version of DeepSeek-V4-Pro, with tower placement and 5-wave flow available.

But overall it feels more like a basic demo: there are only two types of towers, no logs/selling/slowing down/enemy type changes, and the visuals are also relatively...Simple.

DeepSeek-V4-Flash:

DeepSeek-V4-Flash is a well-made single file, capable of placing towers, launching waves, and rewarding kills with money. It also includes event notifications and a tower attribute panel.

The shortcomings are that there are only two types of towers, the enemy and strategy depth is average, and the interface is visually weak.

GLM 5.1:

GLM 5.1's game structure is very product-oriented: 10 waves, two types of towers, multiple enemy types, speed button, event log, and complete information in the right-side shop.

However, the graphics are rather dark, and the combat feedback is not intuitive; there are few types of towers, and none can be sold or upgraded.

Ranking of this case: Hy3 > GLM 5.1 > DeepSeek-V4-Flash > DeepSeek-V4-Pro

Case 4 Code Debugging

Prompt words:

The following Python code finds the indices of two numbers in an array whose sum equals target, but it has a bug. Please point out the bug, provide the correct code, and add three test cases.

def two_sum(nums, target):

seen = {}

for i, n in enumerate(nums):

If target – n is seen:

return [i, seen[n]]

seen[n] = i

return []

Hy3:

Hy3 standard test cases are clearly explained, pointing out that [2,7,11,15] will result in a KeyError, the code fixes are correct, and the test coverage includes basic, repetitive, and unsolvable cases.

The statement that "only when n == target – n is correct" is slightly absolute. If the original return order is strictly required to be [0,1], [3,3], the original code may also be reversed.

DeepSeek-V4-Pro:

The fixes and tests in DeepSeek-V4-Pro are correct, and all four test cases are fully covered. However, the explanation that seen[n] might be "the index of the current number n, i.e., i itself" is not accurate, because seen[n] = i occurs after the judgment, and usually has not yet been written.

DeepSeek-V4-Flash:

DeepSeek-V4-Flash is clearly flawed, misjudging the problem as "return order is reversed," yet it still uses `seen[n]` instead of `seen[target – n]`. The standard test case [2,7,11,15], target=9 still results in a KeyError, and even the provided tests fail to run.

GLM 5.1:

GLM 5.1 accurately identifies the core issue: it checks if `target – n` is in `seen`, so the return value must be `seen[target – n]`. It also provides a more rigorous explanation that `seen[n]` might encounter a KeyError or a duplicate value. The fixed code and tests are correct.

Ranking of this case: GLM 5.1 > Hy3 > DeepSeek-V4-Pro > DeepSeek-V4-Flash

Case 5: Chinese Writing

Prompt words:

Please rewrite the following paragraph into a more natural WeChat official account post, instead of making it look like... AI Written.

original:

existartificialintelligentfastIn today's development,Agent It's no longer just a technological concept; it's becoming a crucial tool for transforming productivity. For ordinary people, understanding... Agent It is very important because it can not only improve work efficiency, but also help us rethink the relationship between humans and machines.

Hy3:

Hy3 has a strong sense of peer review and can accurately point out problems in the original text. AI The source of the melody, and the rewriting, also have the appeal of the beginning of the public account.

DeepSeek-V4-Pro:

DeepSeek-V4-Pro is clean, concise, direct, and not bloated. However, it contains relatively little information. Agent When written as a typical efficiency tool, it lacks key elements such as "proactive execution, multi-step tasks, and changes in human-machine division of labor."

DeepSeek-V4-Flash:

While DeepSeek-V4-Flash uses more colloquial language than the original, it still retains a clear template: "Not just…", "Really…", "Double the work efficiency", "Rethink it." The question of who should listen to whom between humans and machines also has a clickbait feel to it.

GLM 5.1:

GLM 5.1 is specific and scenario-based, going beyond simply "improving efficiency" to describe tasks like booking flights, organizing documents, monitoring progress, generating reports, sending emails, and archiving data, making it easy for readers to understand. Agent The sentence structure is clear and the tone is natural, without too many template sentences. A minor issue is the repeated use of "rather than...it's better to...".

Ranking of this case: GLM 5.1 > Hy3 > DeepSeek-V4-Pro > DeepSeek-V4-Flash

case 6 Constraint Compliance

Prompt wordsPlease output in JSON format, not Markdown or explanation.

Task: Give a AI Design 5 Chinese advertising slogans for programming tools.

constraint:

• 1. Each entry shall not exceed 16 Chinese characters.
• 2. The words "efficiency" and "efficiency" should not appear.intelligentThe three words: “future”.
• 3. Each item must contain a verb.
• 4. The JSON field is "slogans", and its value is an array of strings.

Hy3:

Hy3's JSON format is correct, with exactly 5 fields, all in short Chinese sentences, without English, spaces, or unnecessary explanations.

DeepSeek-V4-Pro:

DeepSeek-V4-Pro's JSON is compliant and the wording is more vivid, but some of the lines are a bit long, such as: just state your requirements, the code itself is long, just throw your ideas in.

DeepSeek-V4-Flash:

The DeepSeek-V4-Flash format is fine, but the constraints are the weakest. It uses a bug...AI This kind of English; bug fixed and immediately deployed, one person per account.AIIt's a bit of an exaggeration to say that it supports the entire stack, like a promotional slogan.

GLM 5.1:

GLM 5.1 was also very stable, with all five bars being short and maintaining a consistent rhythm. However, it used the English word "bug".

This case ranking: Hy3 > DeepSeek-V4-Pro > GLM 5.1 > DeepSeek-V4-Flashcase 7

Code refactoring

Prompt wordsThe function below has been running online for two years and now needs to support coupons, membership discounts, and regional tax rates. Please design a refactoring solution.

function calcTotal(items, user) {

Let total = 0

for (const item of items) {

total += item.price * item.count

}

if (user.vip) total *= 0.9

total *= 1.06

return Math.round(total * 100) / 100

}

Require:

• 1. Avoid over-design.
• 2. Retain testability.
• 3. Provide the TypeScript code.
• 4. Provide a sample unit test.
• 5. Explain which logic should be configured.

Hy3:

Hy3 clearly explains the separation of pure functions, configuration objects, and the boundaries of tax rates/member discounts/coupons, making engineering judgments reliable.

DeepSeek-V4-Pro:

DeepSeek-V4-Pro boasts solid implementation, extensive test coverage, and zero-dependency operation. The drawback is the change from "User" to "membership," which could pose a compatibility risk if the original production version used the "VIP" field.

DeepSeek-V4-Flash:

DeepSeek-V4-Flash has the right approach: constant configuration, fixed computation order, and no strategy pattern.

However, it has the most implementation problems: coupons take precedence over member discounts, and member discounts are deducted on a subtotal basis, which can easily lead to discrepancies between the coupon combination and business intuition; fixed coupons do not have a final safety net, and stacking multiple coupons may result in a negative number, making the whole thing look like a draft.

GLM 5.1:

The GLM 5.1 design is well explained: configuration injection, migration pacing, and what not to over-design are all covered. In particular, the suggestion to first use the old hard-coded configuration for regression testing before connecting to the operations backend is very similar to real engineering processes.

The deduction points are due to minor flaws in the code snippets. pricing.ts does not demonstrate importing types from types.ts, and DiscountPolicy/TaxPolicy are not imported in the tests. Directly copying them may not make the code run.

Case 8 Excelautomaticchange

Prompt words:

I have a sales table with the following fields:

Date, Salesperson, City, Product, Sales Amount, Cost, Customer Type.

Please design an Excel analysis template with the following requirements:

1. automaticCalculate gross profit and gross profit margin.
2. Generate perspective analysis by month, city, and salesperson.
3. Identify orders with a gross profit margin of less than 15%.
4. Generate a management dashboard.
5. Provide the formulas, pivot table field configurations, and chart types to be used.

Hy3:

Hy3 delivers a true .xlsx file, complete with a generation script, formula columns, and conditional formatting.automaticThe summary includes KPIs and three charts. While no actual Excel pivot table was created, descriptions of the pivot fields were provided.

DeepSeek-V4-Pro:

DeepSeek-V4-Pro has a clear structure, a complete 6-sheet design, and its KPI formulas and FILTER low-margin list comparison are practical. The gross profit margin calculation in the pivot table using the "gross profit / sales revenue" field is very precise.

DeepSeek-V4-Flash:

The DeepSeek-V4-Flash solution has a basic framework, but there are many issues with the details of the formulas: the low-margin marker does not prevent division by zero, COUNTIF(I:I,<0.15) is missing quotes, and the profit margin suggestions in the pivot table are not accurate enough.

GLM 5.1:

GLM 5.1 closely resembles the actual Excel setup process. Features like super tables, pivot tables, low-profit alerts, slicers, timelines, and thresholds can all be placed in the settings cell.practicalHowever, some formulas are written in a way that focuses on describing properties, such as = total gross profit / total sales.

Ranking of this case: Hy3 > GLM 5.1 > DeepSeek-V4-Pro > DeepSeek-V4-Flash

Case 9 DCF Valuation

Perform a DCF valuation based on the following assumptions.

Free cash flow in 2026 is 12 million, with growth rates of 18%, 15%, 12%, and 8% for the following four years. The perpetual growth rate is 3%, and the WACC is 10%. Net debt is 20 million, and there are 10 million shares outstanding.

1. Present value of explicit period cash flows.
2. Final value.
3. Corporate value.
4. Equity value.
5. Value per share.
6. Perform sensitivity analysis for WACC 9%-11% and perpetual growth 2%-4%.

Hy3:

Hy3's discounting method is correct: the cash flows from 2026 to 2030 are discounted at t=1..5, and the final value at the end of 2030 is also discounted over 5 years. The benchmark of 21.95 yuan per share is correct, and an HTML report and calculation script were actually generated.

DeepSeek-V4-Pro:

DeepSeek-V4-Pro also has a local script, but its core approach is wrong: it omits the 12 million FCF in 2026 and only starts discounting from 2027; the terminal value is also only discounted over 4 years, so the valuation is inflated to 23.14 yuan.

DeepSeek-V4-Flash:

Both DeepSeek-V4-Flash and DeepSeek-V4-Pro make the same DCF (Discounted Cash Flow) error: omitting the 2026 cash flow and having one less year in the futures discount period. However, DeepSeek-V4-Flash is inferior to DeepSeek-V4-Pro in that it lacks local documentation and its explanations are shorter.

GLM 5.1:

GLM 5.1's calculations and definitions are rigorous, clearly stating the valuation date, year-end cash flow agreement, and 5-year discounted future value. The 5x5 sensitivity analysis is also complete. The conclusion's future value percentage and cross-validation highlights demonstrate financial modeling awareness; unfortunately, no local delivery documentation was generated.

In this case, the ranking is: Hy3 > GLM 5.1 > DeepSeek-V4-Pro > DeepSeek-V4-Flash

Judging from the test results, Hy3's advertising is quite honest. It does outperform GLM 5.1 overall, and I was quite surprised that it outperformed DeepSeek-V4-Pro.

Hy3 is used in front-end development, tower defense, constraint compliance, code refactoring, and Excel. automaticThey achieved first place in the questions on transformation and DCF, and their strength lies especially in their ability to deliver: they can generate documents, run web pages, and break down requirements into real deliverables.

GLM 5.1 excels in code debugging, Chinese writing, and DCF terminology—questions that require judgment. It provides rigorous explanations and demonstrates a strong understanding of financial modeling.

DeepSeek V4 Pro is more like a solid, engineer-like developer. It has good structured code capabilities, and it actually managed to run 20 tests successfully during code refactoring.

DeepSeek V4 Flash is suitable for lightweight tasks. Single-file pages, small demos, etc.fastThe drafts are okay, but they fall short when it comes to rigorous reasoning, adherence to constraints, and complex business terminology.

After testing, my impression of the Hy3 is quite clear: it's really usable for daily work.

Of course, I should clarify first that I used WorkBuddy for this test. It's Tencent's own platform. Agent Using Tencent's own model might offer some compatibility advantages. But regardless, the final results speak for themselves: the page runs, the file can be submitted, and complex tasks can be broken down, so the overall completion rate is indeed quite good.

More importantly, using Hy3 in WorkBuddy is currently free.freeIt takes two weeks. For those who frequently run code, build web pages, and write analytics, this is essentially equivalent to saving half a month's token fees. If you later integrate an API, the price isn't expensive, even slightly lower than DeepSeek-V4-Pro.

Another thing I'm really looking forward to is that Hy3 has already integrated with many Tencent products, such as Yuanbao and WeChat Reading. These are also tools I use regularly.

I also saw that Hy3's computing power consumption peaked around 10 AM on July 8th, and there were queues afterwards, with the queue rate exceeding 50% at one point in the afternoon. This shows one thing: I'm probably not the only one who thinks it's good.

After this test, I will add the Hy3 to my daily work backup list, and I hope that domestically made models will continue to improve.

Original link:Tencent Hy3 is out, let me see if there's any hype in the promotional materials.