Qwen3.8-Max

According to media reports, Moonshot AI's Kimi K3 and Alibaba's Qwen3.8-Max exceed the 2-trillion-parameter threshold, but the draft itself does not name any specific model.

In the command injection Java case, four of the six models passed the mount argument to ProcessBuilder as a list, while Qwen3.8-Max and Kimi-K3 built strings for sh -c, causing the shell to interpret user input as commands.

The second task required generating a city night scene containing a Ferris wheel, a moon blocked by the wheel's spokes, and a complete hierarchy of city lights; in Qwen3.8-Max's output the Ferris wheel structure was clear, the cabins and gears were correctly linked, and the city night scene had a complete hierarchy, but there was one visual error where the moon was occluded by the spokes.

The third task required running Tetris inside the shell of a retro handheld console, testing logical rigor, UI aesthetics, the animation state machine, and the Game Boy-style shell, buttons, and screen colors; Qwen3.8-Max completed the block falling, stacking, and line-clearing logic with no errors and no clipping, and fully reproduced the shell, buttons, and screen colors.

Qwen3.8-Max has a total parameter count of 950B, and its activated context supports extremely long text.

Qwen3.8-Max is competent in all three areas of Chinese writing, 3D modeling, and game logic, and its shortcoming lies mainly in efficiency rather than the ceiling of its capability.

In this hands-on test, when Qwen3.8-Max handled highly complex, multi-constraint tasks, its reasoning and generation chain was very long, taking about three hours to complete four tasks, indicating that it still has room for optimization at the efficiency level before it is "ready to use at hand."

The fourth task was a 3D racing scene, "Sunset Circuit," requiring a 3, 2, 1, Go countdown, a speedometer, lap times, a HUD, and multiple cars on the same track; Qwen3.8-Max produced a complete, playable track, and the car could recover on its own after skidding in a curve.

The first task required writing ten sentences, with the first sentence of three characters, the second of four characters, increasing in turn up to the tenth sentence of twelve characters; Qwen3.8-Max completed the task, with the character counts and the increasing structure all correct.

After the researchers added a sentence to four models requiring them to implement "ensuring security while meeting functional requirements," GPT-5.6-Sol's functional + security pass rate rose from 24.7% to 46.2%, GLM-5.3 from 21.3% to 38.2%, Qwen3.8-Max from 22.0% to 37.6%, and DeepSeek-V4-Pro from 18.0% to 19.1%.

A Bilibili blogger tested Qwen3.8-Max with four tasks on the "Qianwen Office" platform, and it passed all four, with the whole process taking nearly three hours.