Alibaba has previewed Qwen3.8-Max-Preview, a 2.4 trillion-parameter artificial intelligence model designed to process text, images, video and documents. The company is positioning the system as its most capable Qwen model yet, but it has not released the benchmark results, technical documentation or independent evaluations needed to assess that claim.
The model was presented at the World Artificial Intelligence Conference in Shanghai. Alibaba’s Qwen team described it as ranking behind only Anthropic’s Claude Fable 5 among frontier AI systems. That comparison remains an internal assertion: no independent leaderboard has evaluated Qwen3.8, and Alibaba provided no scores or task-level results showing how it compares with Anthropic’s model or other leading systems.
Qwen3.8 is the first model in Alibaba’s Qwen family with more than 1 trillion parameters that can handle multiple types of input, including visual media and documents, in addition to text. Parameter count is one measure of a model’s scale, but it does not by itself establish performance. Architecture, training data, inference methods and the number of parameters activated for each request can all materially affect a system’s capabilities, efficiency and operating costs.
Several important technical details are absent from the preview. Alibaba has not issued a model card, disclosed the model’s activated-parameter count or published information about performance across areas such as reasoning, coding, multimodal understanding and tool use. It has also not provided a direct breakdown against Qwen3.7-Max, the previous flagship released in May.
The lack of evaluation data differs from the rollout of Qwen3.7-Max, which arrived with published test results. That earlier model recorded 56.6 on the Artificial Analysis Intelligence Index. Without equivalent measurements for Qwen3.8, developers and enterprise customers cannot yet determine whether the substantially larger parameter count translates into better performance.
Alibaba said Qwen3.8 is still evolving and plans to make open weights available, though it did not announce a release date or licensing terms. Those details will determine how freely developers can download, modify and deploy the model. Open-weight releases have been an important part of Qwen’s growth, enabling researchers and companies to adapt models for their own applications and infrastructure.
The preview can currently be accessed through Alibaba’s Token Plan subscription and its Qoder and QoderWork development platforms. During the trial period, usage is priced at 10% of the standard rate. Alibaba has not indicated how long that discount will remain in effect.
Qwen3.8 arrives as Chinese AI developers move into the multi-trillion-parameter range. Beijing-based Moonshot AI introduced Kimi K3, a 2.8 trillion-parameter model, three days before Alibaba’s announcement. The closely timed releases illustrate the increasing emphasis on very large systems among Chinese laboratories, even as developers across the industry also pursue techniques intended to reduce the amount of computation used for each query.
Alibaba has made Qwen a core part of its broader AI strategy, pairing a growing model catalog with cloud services and custom computing hardware. Its recurring open-weight releases have also helped the family attract a substantial developer community. Whether Qwen3.8 extends that momentum will depend on the promised weights, their license and evidence that the model can deliver competitive results efficiently.
For now, Alibaba’s comparison with Claude Fable 5 cannot be independently checked. Publication of a model card, reproducible benchmark results and open weights would give researchers and customers their first opportunity to test the ranking and compare Qwen3.8 directly with its predecessor and competing frontier models.
Sources: Anthropic