跪拜 Guibai
← Back to the summary

Tencent's Hy3 Model and WorkBuddy Agent Run a Five-Case Gauntlet

Hello everyone, I'm Second Brother.

People in the comments have been asking me to test WorkBuddy, so here it is today, together with Tencent's newly released Hunyuan Hy3 model.

I believe that with the official release of Hunyuan Hy3, WorkBuddy will become the preferred desktop Agent for more and more people.

Let's start with the parameters.

Hy3 is a Mixture of Experts (MoE) model that integrates fast and slow thinking, with 295B total parameters, 21B activated parameters, and a maximum context length of 256K.

Internally, it has 192 routed experts, activating the top-8 each time, plus one permanently active shared expert. The reasoning intensity is divided into three levels—no_think, low, high—allowing you to decide whether to engage deep thinking based on task complexity.

Let's give it a quick try: use WorkBuddy + Hy3 to generate a poster about MoE.

The prompt was very simple.

@Visual Poster Design a poster for the release of Hunyuan Hy3. The parameter information is here: https://mp.weixin.qq.com/s/X2x1GF09bFbTzc3M1981BQ?scene=1&click_id=1351723181

Open WorkBuddy, select the Hy3 model, choose the Visual Poster skill, and enter the prompt.

WorkBuddy first reads the content from the WeChat public account, extracting key information such as hallucination rate 12.5%→5.4%, task resolution rate 72%→90%, time consumption -34%, Token efficiency for documents -47.4%/PPT -49.0%, etc.

Then it prepares the canvas and starts designing.

During the process, it performs necessary deep thinking and tool calls.

After completing the drawing, it also actively verifies and fixes visual errors.

Finally, it saves the file.

I directly copied and pasted from the canvas area. As you can see, all the parameter information in the poster is correct.

By the way, after the official version of Hy3 is launched, there is a two-week free quota. Those who need massive amounts of Tokens must seize this opportunity.

From my personal experience, Hy3 and WorkBuddy are highly compatible. After testing several cases, its coding ability is very close to Deepseek V4 Pro.

I designed five cases covering four capability dimensions: front-end planning, tool use, long-range execution, and complex reasoning.

The first three correspond to Coding Agent upgrades, and the last two correspond to the office domain of Working Agent.

I'll paste the prompt for each case verbatim so you can reproduce them in WorkBuddy without changing a single word.

01. Neo-Brutalist Front-End

The first case tests planning and front-end. Front-end is also a key upgrade area for Hy3.

I gave WorkBuddy a design prompt in the Neo-Brutalism style and asked it to generate a complete single-page website from scratch. This style has been trending in the design circle recently, characterized by thick strokes, hard shadows, and highly saturated candy-colored blocks, looking like an interface built from colorful building blocks.

The prompt is as follows.

Design a single-page website in the Neo-Brutalism style.

Colors: Use highly saturated candy colors for the main color blocks—Lemon Yellow #FFE156, Fluorescent Pink #FF6B9D, Electric Blue #4ECDC4, Mint Green #95E77E. Use cream white #FFF8F0 or light gray for the background. Each functional area gets one main color, with hard cuts between color blocks and no gradient transitions.

Borders and Shadows: All cards and buttons should have a 3px solid black stroke; shadows should be solid black blocks, offset 6px to the bottom right, with no blur, like the hard shadow cast by building blocks. On hover, the shadow shortens to 2px to simulate a pressing-down feel.

Fonts: Use a 900-weight sans-serif for titles and a monospace font for body text. Titles can be intentionally tilted -2° to 3° to create a casual, hand-labeled feel.

Layout: Use an asymmetric grid with cards of varying sizes, some spanning two columns and others just one. Allow slight overlapping between elements (e.g., a sticker pressing on the boundary of two color blocks). Mix rounded and right-angle corners—large color blocks use 16px rounded corners, small labels use 0 rounded corners.

Texture Details: Add a fine noise texture to the background of at least one area; use 2-3 hand-drawn style decorative symbols (asterisks, arrows, wavy lines) as inline SVGs, with colors matching the main color of their area.

Overall Vibe: Like a magazine layout pieced together with colored tape and markers—bold, with attitude, but with clear information hierarchy. No sense of refinement, but raw vitality.

The constraint density in this prompt is very high: 4 specified color values, stroke thickness and shadow offset precise to the pixel, specific parameters for hover interaction, mixed corner radius rules, and inline SVG requirements. Whether it can ingest all these visual specifications at once and produce a unified output directly reflects the model's requirement understanding and planning capabilities.

02. Let Hy3 Draw a Dynamic Business Card

The second case also tests front-end.

My idea was to let Hy3 draw itself—using web animation to demonstrate the token routing process of the MoE model. This task is a bit tricky; the model must first truly understand its own architecture to draw it correctly.

The prompt is as follows.

Help me create a single-page website with the theme "Hy3's Self-Portrait," using animation to demonstrate the token routing process of the MoE model.
Requirements: Place three digital cards at the top of the page, showing 295B total parameters, 21B activated parameters, and 256K context;
In the middle, a routing animation: a token flows in, 8 out of 192 experts light up, and the shared expert stays constantly lit; at the bottom, place a slider that adjusts the number of tokens flowing in per second, with the animation speed changing accordingly;
Dark background, Tencent Blue color scheme;
Implemented as a single HTML, CSS, JS file, no external libraries allowed.

This prompt has five constraints: digital cards, expert lighting logic, slider linkage, color scheme, and single file.

The more constraints, the clearer it becomes whether the model reads all requirements and plans holistically, or processes them sentence by sentence.

For easier demonstration, I had WorkBuddy generate a GIF animation.

And it specifically processed the frame count to meet WeChat public account requirements for GIFs.

03. Let Hy3 Verify Its Own Data

The third case tests tool use, focusing on browser navigation and information extraction accuracy.

Open three pages: GitHub's Tencent-Hunyuan/Hy3 repository, HuggingFace's tencent/Hy3 model card, and OpenRouter's Hy3 page.
Extract information about context length, open-source license, and pricing from the three sources and create a comparison table. If the data from the three sources does not match, list the conflicting items separately and indicate the source. For uncertain information, write "Not Found" directly, do not guess.

The cleverness of this case is that I knew the answer in advance. The data from the three sources already had discrepancies; for example, Hy3 is actually not listed on OpenRouter.

It can be seen that WorkBuddy opens three web pages, extracts key information, and creates a comparison table.

And it clearly tells us:

04. One Sentence to Produce Excel and PPT

The fourth case tests long-range execution, the home turf of the Working Agent.

This folder contains all the Markdown articles from my blog's AI column. Each article starts with a title, date, and tags.
Step one, organize them into an Excel sheet with columns for title, publication date, category, and word count; Step two, count the number of articles published each month and the proportion of each category, and generate charts; Step three, based on the categorization results, produce an "AI Agent Learning Path" PPT within 10 pages, with the data statistics chart on the last page.
After completing all steps, self-check whether the data in Excel matches the original files, then hand it over to me.

A three-step task, involving cross-file reading, spreadsheets, charts, and PPT—four tools—plus a final self-check requirement.

The key to long-range execution is not that every step is amazing, but that later steps remember the data from previous steps and can self-discover errors before delivery.

I've run this type of task with other Agent tools more than once, and the two most common failure points are:

One, word count statistics—the model is too lazy to actually count and just estimates a rough number to fill in; two, the statistical charts in the PPT don't match the raw data in Excel, with each side calculating independently.

So the last sentence in the prompt, "self-check before handing it over," is not just a polite phrase.

It's said that Hy3 can stably support Agent workflows of nearly 500 steps, which is really, really strong.

Let's look at the final output.

To generate a more aesthetically pleasing PPT, let's install Zang Shifu's PPT Skill here.

Install this https://github.com/op7418/guizang-ppt-skill

Okay, let's generate a Swiss Style PPT.

This task was quite heavy, and context compression was performed several times during the process.

OK, the task is done. Let's see the results.

I really didn't expect that I published 45 pieces of content in March.

Long-range execution tests not just single-step capability, but context retention.

Hy3 maintained data consistency throughout this case.

This is actually a typical task for a Working Agent: first read a bunch of long texts, then extract structured data, and finally deliver results across multiple tools like Excel, charts, and PPT.

05. Let Hy3 Tackle a Complex Reasoning Problem

The last case tests complex reasoning.

Agentic Reasoning is the third major upgrade direction for the official version of Hy3. Let's pose a serial problem requiring mathematical derivation + coding + execution + visualization.

The problem is the classic derangement problem.

100 people stand in a row, each with a numbered hat on their head. The hats are randomly shuffled and distributed. What is the probability that at least one person gets their own numbered hat?
Requirements:
1. First, provide the derivation process for the exact mathematical solution
2. Write a Monte Carlo simulation in Python, running 1 million times
3. Draw a convergence curve graph, with the x-axis as the number of simulations and the y-axis as the probability estimate, also marking the horizontal line for the exact mathematical solution
4. Compare the error between the simulation result and the exact solution

Why this problem?

Because it simultaneously examines the two key upgrade directions of the official Hy3 version: deep reasoning (the derivation of the inclusion-exclusion principle is not a one-step jump to the answer, requiring multi-step logical deduction) and tool use (Python coding + execution + Matplotlib plotting).

Five interlocking steps: problem understanding → mathematical derivation → code writing → simulation execution → visualization comparison.

First, look at the mathematical derivation.

The key is whether Hy3's derivation process is complete.

Did it write out the inclusion-exclusion formula? Did it explain why 1/e can be used as an approximation when n is sufficiently large? Did it calculate the final value correctly? If it directly jumps to "the answer is approximately 0.63" without a derivation process, that's not reasoning.

Final verification: After simulating 1 million times, the error between the probability estimate and the exact mathematical solution of 0.63212 should be at the thousandths level (within ±0.001).

By the way, here's how this capability is used in daily work.

Data analysis, quantitative modeling, and A/B test result validation all require the serial capability of "first derive the formula, then run the data, then draw the chart."

ending

After running five cases, let me share my overall impression.

Tool use checks whether it can open web pages, find information, and cite sources; planning ability checks whether it can break down multi-constraint requirements into executable steps; long-range execution checks whether it loses state across multi-step tasks; complex reasoning checks whether it can first derive, then write code, then execute and verify.

The performance of WorkBuddy+Hy3 fully met expectations.

And Hy3 in WorkBuddy is now free to try for two weeks. Everyone can seize this window period to run some tasks you've wanted to do before but never got around to.

See you next time.