Alibaba Qwen 3.8: 2.4T-Parameter Model Claims 'Second Only to Fable 5', Open Weights Promised

Alibaba’s Qwen team announced its latest large language model, Qwen 3.8, on July 19, 2026¹. The model has 2.4 trillion total parameters, and the team describes it as “one of the most powerful models available today,” on par with leading frontier models and “second only to Fable 5”¹. The announcement also promises that the model is “going open-weight soon”¹.

A preview version called Qwen3.8-Max-Preview went live the same day on Alibaba’s Token Plan subscription service and the Qoder and QoderWork developer tools¹. During the preview period, the model is reportedly offered at 10 percent of the standard price².

What Was Announced: A 2.4-Trillion-Parameter “Multimodal” Model

The announcement came as a post on the Qwen team’s official X account¹. The post highlights the 2.4 trillion total parameter count and says the model is “continuously evolving”¹. According to reports, Qwen 3.8 is the team’s first multimodal model above the 1 trillion-parameter mark, capable of processing images, videos, and documents². The team expects it to outperform the previous-generation Qwen 3.7-Max, particularly in coding and complex productivity tasks².

Token Plan, the main access channel, reportedly offers subscription tiers ranging from $6 to $68 with API access through OpenAI- and Anthropic-compatible protocols³. Being able to try a new model by simply switching endpoints in existing tools and SDKs has become a common distribution pattern among recent Chinese models.

A Preview With Many Blanks Left Unfilled

At the time of the announcement, benchmark results, license details, a model card, and the timing of the open-weight release were all unpublished². The “second only to Fable 5” positioning is based on the company’s own internal evaluations, with no third-party verification yet³. Architectural details—such as whether the model uses a MoE (Mixture of Experts) design and how many parameters are active—have not been disclosed either.

Starting with a proprietary release and opening the weights later mirrors the strategy Alibaba used for the previous Qwen 3.6, according to reports³. For now, the open-weight announcement should be read as exactly that: a promise, not a release.

The Context: China’s Trillion-Parameter Open-Model Race

The announcement landed just three days after Moonshot AI unveiled Kimi K3, a model with 2.8 trillion total parameters and a pledge to release its weights by July 27, 2026 (see our article on the Kimi K3 launch). Qwen 3.8’s 2.4 trillion parameters put it in the same scale bracket³, showing that China’s open-weight competition has entered a phase defined by trillion-scale parameters plus open-release pledges.

The relationship between the two companies adds an interesting twist. Alibaba reportedly owns about 36% of Moonshot AI³, making this an in-house rivalry in which a major shareholder ships a competing model. Kimi K3 has also promised open weights, but access currently runs through its chat app and API². The pace of change becomes clear when compared with July 2025, when Moonshot AI open-sourced the 1-trillion-parameter Kimi K2. In just one year, the field has moved from 1 trillion parameters to open-release pledges for models more than twice that size.

Behind the race is the broader US-China AI decoupling. Alibaba has been part of moves to reduce reliance on US-made AI tools, such as restrictions on Claude Code use in China, and stronger domestic models directly deepen the pool of alternatives.

Wait for the Weights Before Deciding

For readers weighing AI adoption in their business, there are two takeaways.

First, if the weights are released as promised, the range of models that can be self-hosted on private infrastructure may widen further. With Kimi K3 at 2.8 trillion and Qwen 3.8 at 2.4 trillion both pledging releases, the performance ceiling of open-weight models continues to move up.

Second, the evidence needed for a decision is not yet on the table. With benchmarks, license terms, and the release date unpublished, building adoption plans on the “second only to Fable 5” claim would be premature. Since the preview is priced at 10 percent of the standard rate², a realistic two-step approach is to run small-scale evaluations now if the model looks relevant, and defer any production commitment until the official release and third-party benchmarks arrive.