Alibaba offers 2.1 billion QwenCloud tokens for agent feedback
The Qwen Growth Plan asks developers to submit successful and failed Qwen3.8-Max-Preview agent tasks by August 7.
By Ryan Merket · Published
Why it matters
Alibaba is using token incentives to collect the real task traces and failure cases that agent benchmarks miss, while pulling developers deeper into its QwenCloud distribution channel.
Qwen, Alibaba's large-model operation, launched a developer campaign Tuesday offering up to 2.1 billion QwenCloud usage tokens for examples of Qwen3.8 completing, or failing at, real-world agent tasks.
The Qwen Growth Plan arrives nine days after Alibaba previewed Qwen3.8-Max at the World Artificial Intelligence Conference in Shanghai. Developers have until August 7 to submit task scenarios, prompts, outputs and evaluations. Alibaba is explicitly asking for bad results alongside successful ones, turning a promotional contest into a short, structured evaluation cycle.
That approach fits the technical instincts of Eddie Wu, Alibaba's co-founder and CEO. Wu graduated from Zhejiang University of Technology's information engineering college and served as Alibaba's technology director when the business started in 1999. He later became CTO of Alipay and Taobao before returning to lead Alibaba in September 2023. Under Wu, Alibaba has concentrated its strategy around commerce and AI plus cloud computing.
The campaign gives Wu's model organization something benchmark announcements cannot supply on their own: detailed records of how developers frame tasks, where an agent loses track of a workflow and which outputs people consider useful enough to deploy.
A contest built as an evaluation loop
Alibaba divided the awards into three tiers. Ten featured entries will receive 100 million tokens each and may appear in an official promotional video. Fifty excellence awards carry 20 million tokens each, including awards for failures that help identify problems. The first 10 eligible entries receive participation awards of 10 million tokens each.
Across the 70 awards, Alibaba is offering 2.1 billion tokens. Those prizes encourage developers to run longer and more complicated workflows while keeping the spending inside Alibaba's QwenCloud service. They also give Qwen a supply of polished demonstrations for marketing and difficult cases for model evaluation.
Submissions go to QwenOfficial@alibabacloud.com with the subject line "#QwenGrowthPlan + Task Description." Alibaba asks entrants to provide the agent scenario, prompt, Qwen's actual performance, an evaluation or suggestions and contact information. Screenshots, code and compressed files are accepted.
The low-friction submission process has a practical advantage: developers can report failures without converting them into a formal benchmark, reproducible repository or public bug report. Alibaba can then sort those examples by task type and severity before deciding which behaviors warrant post-training work.
Qwen3.8 is still a preview product
QwenCloud's model documentation lists qwen3.8-max-preview as a preview model available through Token Plan, with a 1 million-token context window, thinking support, function calling and built-in tools. Separate Token Plan documentation says Alibaba may improve the model during the preview period and later take it offline or replace it with a production release.
Qwen3.8 can call web search, a code interpreter, web scraping, reverse image search and text-to-image search through QwenCloud's hosted tool system. Those capabilities expand the number of places an agent can fail. A useful answer depends on planning, tool selection, retrieval, execution and final synthesis, rather than a single model response.
Alibaba's WAIC summary described Qwen3.8-Max-Preview as a 2.4 trillion-parameter model and said initial tests placed it behind only Anthropic's Claude Fable 5. Those are Alibaba's claims. The current preview is distributed through Token Plan, Qoder and QoderWork.
The Growth Plan gives Alibaba a way to build task-level evidence around those claims before the production model arrives. It also shifts the conversation toward repeatable work: coding across a repository, researching from multiple sources, processing documents or completing jobs that require several tools.
That is a tougher standard than leaderboard positioning. After Alibaba's WAIC claims put Qwen3.8-Max in direct comparison with other frontier systems, the preview has to convert parity talk into developer trust.
Developer feedback serves Alibaba's cloud strategy
The campaign is small beside Alibaba's infrastructure spending, though it serves the same goal. In February 2025, Alibaba committed at least RMB 380 billion, about $53 billion at the time of the announcement, to AI and cloud infrastructure over three years. Wu called AI a "once-in-a-generation" opportunity and identified cloud computing as Alibaba's clearest route to AI revenue.
Alibaba can fund model development internally, distribute Qwen through its cloud products and subsidize experimentation with token awards. The Growth Plan connects those assets: developers receive compute, Alibaba receives usage and feedback, and the strongest examples become sales material for the model.
Wu has spent his career building technology inside businesses with existing distribution, from Taobao to Alipay. Qwen follows the same pattern. Alibaba does not need to launch a separate model startup or raise outside capital. It can place Qwen inside coding tools, cloud subscriptions and its own internet products, then use campaigns like this one to find the tasks that deserve additional engineering.
The honest inclusion of failed cases is the most useful part of the program. Agent models often look capable in controlled demonstrations and break when a task stretches across tools, files and changing instructions. Qwen3.8 will earn developer confidence when Alibaba can turn those failures into measurable improvements in the production release.