A hands-on article published on Juejin offers a concrete verdict: Qwen4 has indeed improved over its predecessor, but it takes roughly 1–2 hours on several complex tasks without delivering a clearly superior result. The author says it offers worse value than faster domestic rivals. This matters because competition among Chinese foundation models is shifting from parameter counts and marketing toward reliably delivering results.
The author tested tasks including a Cyberpunk-style version of Along the River During the Qingming Festival, game replicas, a mechanical-watch webpage and controller interactions. According to the article, some outputs are already “viewable,” but spatial relationships, logical control and visual fidelity still fall short. Game-replica completion reached roughly 70%–80%, while several complex tasks took about 1 hour 38 minutes, 1 hour 34 minutes and more than two hours. For comparison, the article cites models including HY4 and Sonnet 5.5 as alternatives with better speed or performance.
What this is
Qwen4 is the latest model family from Alibaba’s Tongyi Qianwen. The article focuses on its advanced version for complex tasks. It does not provide the full set of official benchmark results, relying instead on the author’s own tests, consumption records and cross-model observations. It is therefore better treated as an individual experience sample than a comprehensive performance conclusion.
Multimodal models—systems that can understand different forms of information such as text and images while generating webpages, game scenes and other outputs—are moving from “can generate” to “can execute consistently as instructed.” The central contradiction exposed by this test is not that the model is entirely unusable, but that the time and compute required to complete tasks do not match the final quality.
Industry view
We believe the article sends a clear signal: Chinese models are still catching up in foundational capability, but product competition has moved beyond leaderboard rankings toward task-success rates, response speed and total cost. For enterprises, a task that takes two hours and consumes nearly half a month’s quota may affect actual deployment more than the model’s sticker price.
The risks are also clear. First, the test sample is limited, and task types and prompts can affect results, so it cannot establish that Qwen4 as a whole is lagging. Second, the article does not fully develop its comparisons with some competing models, and official benchmark data remain unpublished. More importantly, the author uses overtly emotional demonstration cases, and real development conditions differ from evaluation environments. The conclusion that its “flagship capabilities are insufficient” therefore still requires more reproducible testing.
The competitive landscape is also changing. A large-model arena leaderboard cited in the article shows that the top 10 is dominated by models from Anthropic, OpenAI and Google. The implication is that Chinese vendors’ previous pace of catching up is slowing. If they cannot improve speed, reliability and cost simultaneously, model launches will increasingly struggle to convert into enterprise orders.
Impact on regular people
For enterprise IT: Organizations buying models should look beyond the price of a single call and also account for rework, human validation and waiting time. If a flagship model’s advantages are inconsistent, a cheaper, faster model combined with workflow checks may be more economical than deploying a premium model directly.
For individual professionals: Wait times for complex demos, webpage generation and similar tasks may increase significantly. Before paying, users should test models with their own real tasks rather than relying solely on vendor leaderboards or treating one article’s experience as a universal conclusion.
For the consumer market: As models pursue longer execution times and higher per-task consumption, applications may introduce more expensive long-task packages. What will ultimately determine user choice is still whether the results are accurate and whether the process saves effort.