跪拜 Guibai
← Back to the summary

Tencent's Free Hy4 Flash Built a Game, Opened Its Own Browser, and Refused to Trust Its QA Pass

A no-water free model agent capability test: Hy4 flash (Tencent, free until the end of November) paired with WorkBuddy. Throw a game development task with acceptance criteria in one sentence, and see how far it can go autonomously. Several engineering points worth watching: First, when it receives a vague requirement ("set the numbers yourself but make them reasonable"), it doesn't just fill in values—it first establishes design principles—forcing the "crazy order-taking" path to fail, and guiding players to use learning AI/writing to raise long-term unit prices. Second, it defines "done" as "open the browser and click through it yourself," proactively probing whether the environment has browser automation tools, and writing its own check program covering 12 slots / 5 events / dual endings. Third, it doesn't blindly trust assertions: even when the program reports all passed, it still takes screenshots and visually verifies layout and Chinese encoding. Fourth, before changing requirements, it first does a reachability verification to avoid producing a version that "can't win at all." This article reconstructs the six steps, with the full prompt, acceptance checklist, and a scenario-based judgment on "can a free model be the main workhorse."

I'm Xiao Hu, bachelor's and master's from Zhejiang University, a breaker of action plans, author of "AI Super Individual: The Complete Guide to a One-Person Company," focused on AI training, super-individual creation, and company efficiency coaching.

Yesterday Tencent released a new model called Hy4 flash, free until the end of November.

Hello everyone, I'm Xiao Hu.

I've seen plenty of free things. Whether it's a horse or a mule, you have to take it for a walk. My judgment: first look at actual latency and throughput.

According to Tencent, this model is all about being fast and cheap. It can handle research, document processing, PPTs, and even stare at the screen and click the mouse.

In the official blind test, its satisfaction nearly matched Claude Haiku, which is more than ten times more expensive, but it cost only a fraction.

I threw it a challenge: build a playable game, open the browser to verify it yourself, then revise it.

Step 1: Throw a paragraph, and I stopped watching

The task is the paragraph below. I didn't break it into steps or teach it how to do it.

Help me make a small browser game: "A Day in the Life of a One-Person Company." The protagonist is a one-person company owner. A day has 12 time slots (8 AM to 8 PM). Each slot has 5 options: take an order / write / learn AI / rest / spend time with family. Each action consumes time, adds money or health or skill. You set the numbers yourself but make them reasonable. Goal: earn 100,000 within 30 days, and health must not be below 40. There must be random events: customer haggling, platform throttling, inspiration burst—write one each. The game must be playable to an ending, with a clear settlement for both pass and fail. When you finish, you must open the browser yourself, play it through to accept it: confirm all 12 slots are there, all 5 actions are clickable, numbers change, random events trigger, and you can reach an ending. If you find problems, fix them yourself, and only hand it over when it's playable. Finally, make one revision: change 30 days to 15 days, raise the earning target to 150,000, and make it actually take effect in the game—don't just tell me it's done. Show me screenshots along the way. If information is incomplete, mark it and ask me; don't make things up.

After throwing it, I didn't watch. It took the job, started "deep thinking," and that wait was nearly two minutes.

A bit slow, right? Excusable—everyone is milking free stuff, so the servers are bound to be crowded. Once they start charging and the crowd thins, speed will naturally return.

Step 2: It first thought about "trade-offs," not code

I thought it would start writing code right away. It didn't.

It first said: "I need to figure out a problem first: in a '12 slots × 30 days' game, the real difficulty isn't filling in numbers, but making players choose between 'earning money' and 'staying alive'—if you can win just by taking orders, the game is boring. I'll design the balance along this line."

With just that one line, I knew it wasn't just taking the task and doing it; it first thought about what a decent result would look like.

Then it built a checklist, wrote it into a web file, and double-clicking runs it.

My line "set the numbers yourself but make them reasonable" was vague, and it handled it: "For the numbers, I'll make the 'crazy order-taking' path fail (health can't support it), forcing players to use learning AI/writing to raise long-term unit prices, and use rest to recover health."

Step 3: It checked on its own whether it could open a browser

This is the part that exceeded my expectations the most.

After writing the game file, it didn't say "done." Instead, it said: "The game file is in place. Now comes the critical part—I need to actually run it and play through it, not just look at the code and feel good.

First, let me probe whether this machine can drive a browser for automated clicking and screenshots."

Read that line: it didn't wait for me to prompt; it was going to actually run it.

It laid out the probe results: this machine has the tools to automatically open a browser and click pages, and it can click through the real browser without relying on imagination.

Then it wrote its own check program, clicking through all 12 slots, 5 actions, numbers, three event types, and both pass and fail endings.

Step 4: The check said all passed—it didn't believe it

It did click through a dozen-plus items, all passed. At this point, it was within my expectations.

Then it said something that made me sit up: "But I don't want to trust only assertions—visuals are for real humans. I need to look at screenshots myself to confirm the layout hasn't collapsed and Chinese didn't garble."

Its own program said all passed, but it insisted on seeing it with its own eyes. That conscientiousness is worth more than two words "all passed."

Then it actually took screenshots and looked at them one by one.

Step 5: It zeroed in on whether random events actually popped up

The acceptance wasn't over yet. It added another line.

After looking, it said: "The event log shows all three event types left orange-highlighted records, and the numbers moved too." It really looked at the screen, not just the program's two words "passed."

Step 6: Before revising, it asked "after the change, is it still passable?"

The last step was written in my task: change from "30 days, earn 100,000" to "15 days, earn 150,000."

I thought it would just swap the numbers and say "done."

It actually thought one layer deeper.

I never taught it that. Changing to a version that can't win at all means the change is meaningless—it thought of that itself.

After changing, it ran another verification round, all 15 items passed, and it specifically checked that the interface text actually changed.

The game it played looks like this

Let me be clear so you don't misunderstand: it's a web page, not the kind with little characters running around and animated scenes.

But it's playable. Click the 12 slots, and money, health, and skill change on the spot; haggling, throttling, and inspiration pop up in the middle; click through to the end and it settles: if you earned enough, you get a pass screen; if not, a fail screen, and you can play again.

These two screenshots are the endings it played out itself:

The checklist it handed in at the end looks like this:

Final words

One thing to say: can Hy4 flash, free until the end of November, be the main workhorse?

My judgment: for this kind of multi-step, self-verifying, mid-task thinking work, you can trust it.

You throw a vague requirement at it, and it sets its own standards, does the work, verifies, and revises itself—you barely need to intervene.

For a small boss juggling multiple things, the savings aren't minutes, it's the peace of mind of not having to watch over it.

Tencent's blind test reached the same conclusion: satisfaction nearly matched the expensive Claude Haiku, the cost per task was a fraction, and the time was the shortest—Tencent tested this itself; I'm marking it as such, not passing it off as an independent review.

But don't treat it as all-purpose. Don't expect it for long, hardcore programming—in the official benchmarks, the hard engineering coding test scored only in the teens, while its big brother Hy4 got forty-something, a big gap.

That's the full hands-on record. That prompt can be reused directly. And the observation methodology—"throw a vague requirement → see if it sets its own standards → see if it verifies its work → verify reachability before changing"—can be applied to any model to quickly map its autonomous boundary.