Alibaba put Qwen3.8-Max-Preview into Qoder and Token Plan with two numbers in front of developers: 2.4 trillion parameters and Credits consumption that can fall to 0.01x.
The model is still a preview, and the public evidence is incomplete. There is no full model card or independent benchmark package yet. The launch therefore reads less like a finished victory lap and more like a field test: let developers run real coding, document and office workflows, then see whether the bill and the results hold up.
The price cut comes first
Qoder's documentation says Qwen3.8-Max-Preview is the newest foundation model in the Qwen family, with 2.4T parameters and improvements over Qwen3.7-Max in coding and professional Cowork tasks.
Qwen3.8-Max-Preview is the latest foundation model in the Qwen family, with 2.4T parameters.
The campaign covers Qoder Desktop, the JetBrains plugin, Qoder CLI, QoderWake, Cloud Agents, Mobile/Web App and QoderWork. It started on July 19, 2026.
The billing terms are unusually specific. During regular hours, the Credits multiplier drops from 0.5x to 0.05x, a 90% discount. From 22:00 to 08:00 UTC+8, it drops to 0.01x, or 98% off. Qoder says the discount affects Credits pricing only, with model quality unchanged.
Big parameters still need receipts
Alibaba Cloud Developer Community materials describe the model as a top-end Qwen preview for text, images, short video and multi-format documents. They list a 128K standard context window and 1M tokens for special scenarios, with access through Bailian Token Plan, Qoder and QoderWork.
Those numbers put the preview in the top tier of Chinese models: 2.4T parameters, long context, multimodal input, coding workflows and a deep off-peak discount. Generation NT noted the missing proof point: Alibaba is positioning the model near the frontier, but detailed benchmarks and a model card are still absent.
The real test is workflow load
The most important part of the launch is the distribution path. Alibaba did not only expose a cloud API. It placed the model inside coding, office and subscription products where users will ask it to read repositories, migrate code, process slides, summarize documents and run repetitive agent tasks.
That creates a harsher test than a launch post. A model can fail by missing one file, breaking one interface, losing a long instruction chain or making a spreadsheet mistake. Cost also compounds because agent workflows can require dozens or hundreds of tool calls.
What to watch
The next checks are concrete: when the open weights arrive, what license they use, whether official benchmarks cover SWE-bench, long-horizon agents, document understanding and multimodal tasks, and whether pricing remains attractive after the preview campaign.
Alibaba is not only saying the model is large. It is pushing a large preview model into daily work surfaces and making the first test cheap enough to run at scale. Developers' overnight migrations, office document batches and billing pages will decide how much of the launch survives contact with production work.
Sources: Qoder documentation, Alibaba Cloud Developer Community, Alibaba Cloud Model Studio, Generation NT, CocoLoop; checked parameter scale, discount scope, access paths, context window, open-weight pledge and missing benchmark evidence.