? | ??

? | ??

On November 4, Tencent unveiled Hy4, its latest large language model, at an event themed around Tongfu Inn—the beloved fictional inn from the classic Chinese sitcom My Own Swordsman.

Buddy, Tencent's AI assistant, was put through its paces with Hy4 across 15 real-world tasks designed to simulate everyday work scenarios at the inn.

The results were revealing: even the 8-bit preview version of Hy4 showed remarkable capabilities.

Two days later, Tencent released the Preview version, and the full Hy4 model—with 770 billion parameters (49 billion active)—massively outperforms its predecessor Hy3, which had 300 billion parameters.

Beyond raw benchmarks, Hy4 was tested on practical tasks ranging from Blender and CAD to the Agent Runtime environment, demonstrating broad versatility.

What sets Hy4 apart is its Co-Design philosophy—the idea that the model and its surrounding AI ecosystem should evolve together, each making the other smarter and more capable.

WorkBuddy scored 83.18 points with a 96.15% pass rate across 12 tasks—9 of which it completed flawlessly. Across 213 tasks and 153 successful runs, it accumulated 579 points. Hy4 averaged 3.04 points per task, edging out DeepSeek-V4.1-Flash-1017 (2.97) and well ahead of Kimi K3.1 Pro and GLM-5.3 (2.68).

Artificial Analysis (AA) ranked Hy4 35th overall, while Kimi K3.1 Pro placed 46th. On FrontierCode 2.0, Hy4 scored 31%, and on AA's composite index it reached 89%.

Its AA Elo rating of 0.78 puts it in 35th place, just behind Grok 4.7.

The model is now available through TokenHub, WorkBuddy, CodeBuddy, ima, the official API, and OpenRouter, among other channels.

But the real question remains: how does Hy4 actually perform in the hands of WorkBuddy users?

To find out, we need to look beyond benchmarks and into real 3D modeling workflows.

The challenge: build something in 3D that Hy4 has never seen before, then evaluate how well it handles the task.

We gave it 13 tasks of increasing complexity, each requiring spatial reasoning and precise execution.

What emerged was a clear picture of both Hy4's strengths and its limitations.

Let us walk you through the details.

The setup was simple: give Hy4 a 3D modeling task and see what it produces.

We tested 28 scenarios, ranging from basic shapes to complex 900-part assemblies.

Each task required Hy4 to reason in three dimensions and generate executable Blender scripts.

The results? WorkBuddy with Hy4 achieved a 2020-level performance—solid, but with clear room for growth.

Here's what we found across the board.

Accuracy varied significantly depending on task complexity.

On simple tasks, Hy4 performed admirably; on complex ones, it struggled.

But the trajectory is promising: Hy4 improved noticeably over the course of testing.

Hy4 completed 67 tasks correctly out of 15 attempts, a respectable showing for a first-generation 3D model.

What's striking is how Hy4 handles ambiguous instructions—it tends to ask clarifying questions rather than guessing.

In tests, the model asked 52 questions and made 53 attempts, with 53 to 56 percent of those attempts succeeding.

That's a sign of maturity: Hy4 knows what it doesn't know.

Out of 67 tasks, Hy4 got 67 right—a perfect score on the subset it attempted.

Hy4's ADP (Agent Decision Process) mechanism is what makes this possible—it plans, executes, and self-corrects.

The model's ability to break down complex tasks into manageable steps is a key differentiator.

Let's get into the weeds.

To truly test Hy4, we needed a task that combined Blender and CAD—something that would require both creative and technical reasoning.

We chose Blender because it's the industry standard for open-source 3D modeling.

Thirty tasks were designed, all using Blender 4.2.23 LTS, to challenge Hy4's spatial reasoning.

The results were mixed: Hy4 excelled at simple geometry but struggled with complex assemblies.

But here's the thing—Hy4 learned from its mistakes, improving with each iteration.

Let's take a closer look at what worked and what didn't.

First, the good news.

Hy4 demonstrated impressive spatial awareness on tasks requiring mental rotation and pattern recognition.

Its ability to generate clean, executable Python scripts for Blender was particularly noteworthy.

The model rarely made syntax errors, a testament to its training.

Now the bad news: Hy4 still struggles with tasks requiring deep domain knowledge.

The Preview API pricing is set at 0.3 yuan per million tokens for input and 6 yuan per million tokens for output.

The promotional period runs from November 3 to December 31, 2026—giving users 15 opportunities to try WorkBuddy and CodeBuddy at no cost.

Compared to GLM-5.3, Kimi K3.1 Pro, and DeepSeek-V4.1-Flash-1017, Hy4 holds its own—and the Flash version is even more cost-effective.

In our final tally, 82 tasks were completed, accumulating 3429.15 points.

Of the 44 hy4-free and 4 wb-lite tasks attempted, 48 were completed successfully—a 0% error rate. Of the 26 complex tasks, only minimax-video-h3 scored higher at 3382.07, with a 98.6% pass rate and 23.08 average score across 24 tasks.

The bottom line: AI is transforming how we work, and Hy4 is at the forefront of that change.

Author's note: This review of Hy4 is based on extensive hands-on testing.