Alibaba Previews 2.4-Trillion-Parameter Qwen3.8-Max Days After Moonshot's Kimi K3

Two days after Moonshot AI open-sourced its 2.8 trillion-parameter Kimi K3, Alibaba used the World AI Conference in Shanghai to preview a rival of its own: Qwen3.8-Max, a model the company says carries 2.4 trillion parameters and marks its first multimodal system to cross the trillion-parameter threshold.

The timing looks deliberate. China's frontier-model race has increasingly played out through rapid-fire announcements, and Alibaba's decision to unveil Qwen3.8 at WAIC, just 48 hours after Kimi K3's launch, reads as a direct response to Moonshot's move. Shuai Bai of the Qwen team framed the release as a milestone for the group, noting that no prior Qwen model had combined multimodal capability with parameter counts above 1 trillion.

Architecturally, Qwen3.8 relies on a Sparse Mixture-of-Experts design, the now-standard approach for scaling large models without proportionally increasing inference cost. Alibaba says the model handles text, images, video, and documents, and that it is compatible with both OpenAI and Anthropic API protocols, a practical detail that could ease integration for developers already building on either ecosystem.

On performance, Alibaba's claims are notable but unverified. The company says Qwen3.8 should surpass its predecessor, Qwen3.7-Max, in coding, full-stack development, data analysis, and general office-productivity tasks. No benchmark scores or a formal model card accompanied the preview, so these comparisons currently rest entirely on Alibaba's own characterization rather than independent testing.

Access and Open Weights Remain Unsettled

Qwen3.8-Max-Preview is already accessible through Alibaba's Token Plan, its Coder product, and QoderWork, all offered at 10% of standard pricing during this preview phase, presumably to encourage early adoption and feedback. Alibaba has also said it plans to release open weights for the model, but as of the July 19 announcement, no release date or licensing terms have been specified. That leaves a gap between the open-weight positioning implied by comparisons to Kimi K3, which launched with open weights immediately, and the more limited preview access Qwen3.8 currently offers.

The parameter count itself deserves a caveat: the 2.4 trillion figure comes from Alibaba and has not been independently confirmed through third-party benchmarking or technical documentation. The same caution applies to the performance claims against Qwen3.7-Max, which remain promotional until backed by published results.

What's clear is the pattern underneath the announcement. With Moonshot's Kimi K3 and Alibaba's Qwen3.8 landing within days of each other, both scaled well past the trillion-parameter mark, Chinese AI labs appear locked into a scaling contest that increasingly plays out in public, model-card details and independent verification notwithstanding.