Alibaba's Qwen team has previewed Qwen3.8-Max-Preview, a new flagship model that the company says has 2.4 trillion parameters and will go open-weight soon.

The preview is available now through Alibaba's Token Plan, Qoder, and QoderWork. Qoder's launch page says Qwen3.8-Max-Preview is the latest foundation model in the Qwen family and claims significant improvements over Qwen3.7-Max in coding, professional productivity, full-stack development, data analysis, and office workflows.

The launch comes just days after Moonshot AI's Kimi K3 announcement, turning July into a fast-moving open-weight race between Chinese AI labs.

Hands-on video

Alibaba has not published a dedicated launch video that we could verify. A same-day hands-on test of Qwen3.8-Max-Preview is already available on YouTube, showing early coding, game-development, and frontend experiments with the preview model.

What Alibaba announced

In its official launch post, Qwen said Qwen3.8 is "launching and going open-weight soon" and described the model as one of the strongest available today, comparable to leading frontier systems and "second only to Fable 5." South China Morning Post reported the same claim and noted that the announcement landed during the World AI Conference in Shanghai.

The important detail is that this is a preview, not the final open-weight release. The model name developers can use now is Qwen3.8-Max-Preview. Qoder says the preview is available automatically to eligible Pro, Pro+, Ultra, Pro Trial, Teams, and Enterprise users after updating to the latest Qoder version and switching the model tier selector.

Qoder is also using aggressive launch pricing. The Qwen3.8-Max-Preview campaign gives eligible users 90% off credit consumption during regular hours, effectively 10x usage, and up to 98% off during off-peak hours from 22:00 to 08:00 Singapore time. Qwen Cloud's Token Plan page also lists Qwen3.8-Max-Preview among the advanced models available through its subscription bundle.

The claims and the caveats

Alibaba is making three big claims:

  • Scale: Qwen3.8-Max-Preview is described as a 2.4-trillion-parameter model.
  • Capability: Qwen says it is comparable to leading frontier AI models and trails only Fable 5.
  • Access: Qwen says the model is going open-weight soon.

Each claim matters, but each still needs follow-through.

There is not yet a full public model card, license, benchmark table, Hugging Face repository, or active-parameter count. MarkTechPost's analysis makes the key technical caveat: if Qwen3.8 is a sparse Mixture-of-Experts system, the headline total parameter count is not the same as the amount of compute active for each token. That active-parameter number will matter more for self-hosting cost than the 2.4T headline.

The open-weight promise also needs a date and license. Alibaba's Max-tier models have often been hosted previews first, so the community will be watching whether Qwen3.8 becomes a downloadable model family, a limited open-weight release, or a preview that stays tied to Alibaba's cloud tools for longer than expected.

Why developers care

Qwen is positioning Qwen3.8-Max-Preview around agentic work, not just chat. Qoder's launch page calls out long-horizon tasks such as full-stack development, data analysis, and office workflows. That lines up with the broader 2026 model race: labs are competing to make models that can operate inside developer tools, write and revise code, analyze files, and keep context across multi-step tasks.

That is also why Qwen3.8 is being compared with Kimi K3. Both launches are trying to show that Chinese labs can ship massive, developer-facing frontier models with open-weight ambitions. Moonshot's Kimi K3 emphasizes long-context coding and multimodal work. Alibaba's Qwen3.8-Max-Preview is arriving as a hosted preview first, with discounted access through its own developer and agent products.

For builders, the practical question is not whether the launch post sounds impressive. It is whether Qwen3.8 can beat their current model on real tasks: repository edits, UI generation, data analysis, long documents, and tool-using agents. Until independent benchmarks and open weights land, the preview is best treated as something to test, not something to trust blindly.

What to watch next

The next updates should answer five questions:

  • When will the open weights ship?
  • What license will Alibaba use?
  • What is the active-parameter count?
  • How does Qwen3.8-Max-Preview score on independent coding and agent benchmarks?
  • Will smaller or distilled Qwen3.8 variants arrive for local developers?

If Alibaba delivers on the open-weight promise, Qwen3.8 could become one of the biggest open-model releases of the year. For now, it is a serious preview with a very loud claim attached: 2.4 trillion parameters, frontier-model ambition, and another sign that China's open AI race is accelerating.