Qwen3.8-Max Preview Tested: Strong Coding but Slow Thinking

Alibaba released Qwen3.8-Max-Preview, a new flagship base model reportedly featuring 2.4 trillion parameters. Open for testing on the official website and the coding app Qoder, it has triggered extensive hands-on tests within the community, though feedback is notably polarized.

Highlights: Coding and Workflows

The model received praise for programming and task execution. @johnseach claimed it wrote 1,500 lines of astrophysics code in one go with zero errors, while @APPSO found its webpage generation extremely fast, capable of producing complex Three.js 3D pages. In Qoder workflow tests, @HeyNayeem noted the model continuously drives tasks to completion with minimal human intervention. Additionally, @青稞AI integrated it to develop complex features like shared wallets, and @Askmasrmod praised its outstanding creative writing capabilities.

Controversies: Slow Thinking and Unmet Expectations

Despite its impressive capabilities, the preview version has clear flaws and divided口碑. @curiousily found the model sometimes gets stuck in a thinking loop, with front-end skills falling short of promotional claims. @Askmasrmod pointed out that its biggest issue is excessively long thinking times, even for simple prompts. Overall, @davidtsong observed sharply divided feedback on X, with some arguing it fails to meet the expected "Fable level." @teortaxesTex candidly stated that Qwen preview versions are typically rough, serving more as scientific references, and estimated its true performance around the GLM 5.2 tier or slightly better.

Testing Suggestions

Addressing the model's instability, @terryyuezhuo (repost) suggested evaluating the API rather than the web UI, as the web chat is merely a free trial portal with inconsistent output quality. @vista8 also reminded users that the web interface is currently best suited for text and simple code tests, though friends reported the model is indeed getting stronger.

2026-07-19 ~ 2026-07-21 · 13 related posts

Full story(17 episodes)→

Primary sources