Qwen3.8, Kimi K3, and GPT-5.6 tested in 'Thunder Jet' game generation

Over the past week, Qwen3.8-Max-Preview and Kimi K3 were unveiled, pushing the parameter scale of China-developed large models beyond 2 trillion.

However, the competition among large models has long moved beyond merely looking at benchmarks or parameters. The best measure of a base model’s capability is its ability to genuinely integrate into everyday workflows.

To find out for sure, Biteye has decided to take the donkey out and see if it’s a horse.

For the same prompt, “Create a Biteye-style Thunder Fighter web game in one sentence,” the Qwen3.8, Kimi K3, and GPT5.6 Sol models produced entirely different results:

• Qwen3.8: Capable of action, and only capable of action
• GPT-5.6 Sol: Looks great visually, like a brand website
• Kimi K3: Most features, most gamified

⚠️ Friendly reminder: Due to computational limitations, Kimi K3 is no longer accepting subscriptions; this test was conducted using WorkBuddy.

https://x.com/i/status/2079865420576927996

🔗：Try Qwen3.8

Opening Qwen3.8, my first impression was: Are you kidding me?

Dark blue background with a logo that is not native to Biteye in the center, and a "Start Game" button. The black text in the title blends into the dark background, as if it had activated隐身 mode in advance.

After entering the game, the basic functions are at least勉强:

• Drag the fighter jet with your mouse
• Auto-fire bullets
• Red triangle enemy aircraft
• Score breakdown
• Collision and termination mechanism

In testing, the fighter jet was able to fire continuously, and the score rose to 858, confirming that Qwen3.8 at least ran successfully.

The problem is that the entire game feels like a newly built Canvas demo: enemy aircraft are triangles, bullets are dots, and the gameplay barely changes. There’s no weapon progression, no item system, and no clear level pacing.

As the Qwen team itself announced, the post-training for Qwen3.8 is not yet complete and is currently being iterated on a daily basis.

Clearly, the art and design teams haven't joined the group yet.

https://x.com/i/status/2079865420576927996

🔗：Try GPT-5.6 Sol version

Opening GPT-5.6 Sol version, the momentum immediately picks up.

On the left is a clear task, in the center is the combat area, and on the right is a visualized radar and telemetry module. The dark background paired with fluorescent cyan, combined with mixed Chinese-English text and Biteye brand elements, creates a striking 007 agent dashboard aesthetic.

Its system is also significantly more robust than Qwen's, for example:

• Wave
• Armor armor
• Overdrive mode
• Combo Strike
• All-time high
• Pause function
• Mouse and keyboard dual operation

In actual testing, the game ran successfully and achieved a score of 922. Instead of simply popping up "Game Over" upon failure, it displays "Fighter Offline," and restarting is labeled "Reconnect." From start to finish, the messaging and visuals maintain a consistent世界观, which is quite fancy.

However, the problem with GPT-5.6 Sol is that it’s too focused on packaging. The information panels on both sides take up a lot of space, squeezing the actual gameplay area into the center. It feels more like a highly polished brand campaign page that happens to include a small game.

Compared to Qwen3.8, the art, branding, and operations teams for GPT-5.6 Sol are all in place, but the game design team clearly slacked off a bit.

https://x.com/i/status/2079865420576927996

🔗：Try Kimi K3

Although the homepage of Kimi K3 isn't as stunning as GPT-5.6 Sol, the moment the game instructions appear, the feel is completely different:

• W: Power Upgrade
• S: StarShield
• B: Bomb
• H: Fixed
• Space: Deploy the bomb
• P: Pause
• M: Mute

After entering the game, you have three lives, a firepower level, a bomb count, a pause button, and a sound toggle.

The other two versions are still slowly figuring out how the plane moves, while Kimi K3 is already considering how players can interact with items after picking them up.

This is the most obvious difference among the three versions: Qwen implemented the shooting functionality, Sol created the visual packaging, and Kimi proactively added items, resources, progression, and active skills.

Of course, its visuals are still not detailed. Enemy aircraft and effects are mostly simple geometric shapes, and some logo assets are applied quite directly. But as a small game, it knows best how to constantly give players something new.

If we treat the three models as new employees, this test would be something like:

Taking Biteye's Thunder Fighter test as an example, it's now not difficult to have a large model write code to create a webpage where "the plane moves and bullets are fired."

But what truly sets the difference is whether, after pre-training on logic, the model proactively designs feedback, levels, items, progression, and visual themes. Large models must not only understand code, but truly understand why players are willing to play one more round.