Alibaba Launches Preview of Qwen 3.8-Max with 2.4 Trillion Parameters Aiming for Second Place Behind Fable 5
At the WAIC in Shanghai, Alibaba responded to Moonshot's Kimi K3 with a preview sold at 10% of the standard price. No model card, no active parameters disclosed.
On July 19, Alibaba introduced the Qwen 3.8-Max-Preview on stage at the World AI Conference (WAIC) in Shanghai. The model boasts a total of 2.4 trillion parameters within a sparse Mixture-of-Experts architecture and is the first multimodal system from the team above 1 trillion, capable of processing text, images, video, and documents. According to Shuai Bai, a developer of Qwen, the new flagship is "second only to Fable 5" among the models the team used as a reference, positioning the Chinese group's product ahead of OpenAI's GPT-5.6 Sol and the Kimi K3, launched by a local rival just three days earlier.
The preview is available through the Token Plan subscription, on Qoder and QoderWork at 10% of the standard price, and Alibaba has promised an open-weight release "soon," although no specific date was provided. What the company did not publish is as relevant as what was announced: there is no benchmark table, model card, or license; and the number of active parameters per token, which defines the actual cost of serving a sparse MoE, was omitted from the announcement. Without this figure, the comparison with Fable 5 and GPT-5.6 remains in the realm of marketing rather than cost engineering.
Preview in Response to Kimi K3
The timing is not coincidental. On July 16, Moonshot AI released the weights of Kimi K3, a 2.8 trillion parameter model that, according to independent benchmarks cited by Simon Willison and Bloomberg, ranked fourth among frontier models, trailing only Claude Fable 5, GPT-5.6 Sol, and exceeding Anthropic's own Opus 4.8. Kimi K3 already operates with a context of 1 million tokens at $3 per million input and $15 for output, a range identical to that of Fable 5. Alibaba needed to respond before the narrative took hold.
Bloomberg reported that Alibaba's stock rose by up to 5.4% in Hong Kong following the announcement, reflecting perceptions that the company continues to compete with Moonshot, DeepSeek, and ByteDance. DeepSeek entered the scene three weeks ago with V4, which introduced peak pricing, charging double during peak hours in Beijing. Conversely, Qwen 3.8-Max took the opposite approach, offering a 90% discount during its preview phase to drive adoption.
Where the Comparison with Fable 5 Should Be Taken with Caution
Fable 5, launched by Anthropic on June 9, has public numbers: 95.5% on SWE-Bench Verified, 80.3% on SWE-Bench Pro, and an Elo score of 1649 at the WebDev Arena. Alibaba has not published the equivalents for Qwen 3.8-Max and only claims that the new model surpasses Qwen 3.7-Max in coding, full-stack development, data analysis, and office workflows. As long as the model card remains unpublished, the assertion of being "second only to Fable 5" remains a manufacturer’s promise, not an auditable result. Analysts from mysummit.school noted the same pattern in previous releases from the Qwen family, where aggressive comparisons were adjusted downward weeks after the publication of the cards.
Implications Beyond China
For CIOs in Europe and America, the practical effect is the same, regardless of the outcome of the technical showdown: the frontier of multimodal models above 1 trillion parameters is no longer an exclusive club limited to Anthropic, OpenAI, and Google. Kimi K3 and Qwen 3.8-Max account for nearly 5.2 trillion parameters released in the same week by two Chinese labs. Platform teams in the U.S., U.K., and India now need to assess the supply chain, data governance, and regulatory constraints of cutting-edge Chinese models, not just Western options. Eesel AI and explainx.ai have already published operational comparisons between Qwen 3.8-Max and Fable 5 aimed at corporate buyers.
In the Latin American market, providers like Falconi and CI&T that integrate models from Alibaba through Alibaba Cloud now have an option near the top of benchmarks at preview costs, shifting the buying argument in comparison to Western hyperscalers. However, those operating workloads sensitive to compliance must remember that Alibaba has not disclosed where the model will be served or what restrictions will apply to corporate prompts, information that the model card will need to address when published.
The preview of Qwen 3.8-Max does not conclude the race against Fable 5. It establishes a parallel race regarding transparency: whoever publishes the auditable card first will set the benchmark by which future Chinese announcements will be measured.