Alibaba Previews Qwen3.8-Max, a 2.4-Trillion-Parameter Multimodal Model With Open Weights to Follow

Alibaba's Qwen team has opened preview access to Qwen3.8-Max, a 2.4-trillion-parameter model the company describes as its most capable system yet, with open weights promised at an unspecified later date. The announcement came through Qwen's official account on X, positioning the release as a multimodal flagship rather than a text-only update.

• Available now in preview through Alibaba's Token Plan, Qoder, and QoderWork.

• Multimodal input for images, video, and documents.

• Open weights "soon," with no date and no Hugging Face model card posted yet.

A trillion-parameter model built to read images, video, and documents. Qwen positions Qwen3.8-Max as the team's first multimodal model above 1 trillion parameters, able to take images, video, and documents as input, according to The Decoder. That input range is the practical detail for production teams: a model that ingests footage and reference documents directly fits review, logging, and analysis workflows a text-only model cannot serve. Alibaba also expects it to outperform the prior Qwen3.7-Max on coding and productivity tasks including full-stack development, data analysis, and office work.

The video-as-input framing tracks with the pitch behind Moonshot's Kimi K3, which we covered as a 2.8-trillion-parameter model released days before. The Decoder reports Alibaba's release is likely aimed at Kimi K3's momentum.

"Second only to Fable 5" is a vendor claim, not a benchmark. Qwen says the model matches leading frontier systems and trails only Fable 5, but the company has published no benchmark results to support the ranking. Alibaba has also not disclosed the active-parameter count or the mixture-of-experts configuration, so the 2.4-trillion figure describes total size, not the compute used per query. Without published numbers or third-party testing, the positioning stands as a company claim.

Preview access runs at a tenth of standard pricing. The Decoder reports the preview is available through Token Plan, Qoder, and QoderWork at 10 percent of the standard price, giving developers a low-cost window to test the model before general availability. Qwen called the preview "continuously evolving," which means the version teams test today may not match the eventual released weights.

Open weights would extend Alibaba's release pattern, but they aren't here yet. Alibaba has repeatedly shipped models to a paid preview first and released open weights later. The same sequence played out with Wan 2.5, which launched as a paid API with open weights promised down the line.

The company's open-source track record is real, from the freely downloadable Qwen Image Edit to a run of open model families across image and video. Qwen3.8-Max stays preview-only for now, and the open-weights claim rests on the announcement alone.

The open-weights promise is the part still missing. The usable facts are narrow: a multimodal model in paid preview, priced low, accepting image, video, and document input, with no independent benchmarks. The open weights that matter most for self-hosting and fine-tuning, the same draw behind Thinking Machines' Inkling and other open releases, remain a promise without a date. Developers can test the preview today; teams planning around open weights should wait for the model card and real numbers.