Alibaba previews Qwen3.8-Max, a 2.4-trillion-parameter multimodal model

Alibaba's Qwen team previewed Qwen3.8-Max-Preview, its new flagship, at the World AI Conference in Shanghai. The specs are frontier-class: a 2.4-trillion-parameter sparse Mixture-of-Experts architecture, native multimodality across text, images, video, and documents, and a 1-million-token context window. Alibaba's benchmark claim is aggressive — 'second only to Anthropic's Claude Fable 5' among frontier models — and it's offered in preview through Alibaba's Token Plan subscription at just 10% of standard pricing, an explicit price-war move.

On July 28, Alibaba escalated the developer push, launching the #QwenGrowthPlan to gather feedback on Qwen3.8's agentic capabilities, inviting users to complete real-world tasks and submit results. The program is a data-and-adoption flywheel aimed at the same enterprise agent workloads that AWS, Google, and xAI are chasing this week.

The strategic frame is the surge of cheap, capable Chinese open and near-open models. Qwen3.8 Max sits alongside Moonshot's Kimi K3 and Zhipu's models in a cohort that Tavily/Perplexity sources describe as 'nearly as intelligent as OpenAI and Anthropic frontier models — but cheaper and open,' making real inroads in the US and echoing the DeepSeek shock of early 2025. Microsoft testing Kimi K3 on Azure to cut $600M in Copilot costs is the clearest sign the pressure is landing on Western incumbents.

Skeptical note: Alibaba's 'second only to Fable 5' claim is self-reported, and preview pricing at 10% of standard is a customer-acquisition tactic that won't reflect production economics. What to watch: independent benchmarks, whether Qwen3.8 Max opens its weights like Kimi K3, and how US export and procurement policy responds to Chinese frontier models gaining share.