On July 19, 2026, Alibaba's Qwen team previewed the Qwen3.8-Max-Preview during WAIC in Shanghai. The model has a total parameter count of 2.4 trillion and was described as a frontier system second only to Fable 5.
Facts
The preview version is now available through the Alibaba Token Plan subscription, priced at 10% of the standard rate. The model supports text, image, video, and document processing, making it Qwen's first multimodal model exceeding 1 trillion parameters. The announcement stated that it outperforms Qwen3.7-Max in coding, full-stack development, data analysis, and office workflows. The model's weights are planned to be released soon.
This release comes just two days after Moonshot AI launched Kimi K3, which has 2.8 trillion parameters and is currently the world's largest open-weight model. The timing of the two labs' announcements overlaps, both choosing to present during WAIC.
Mechanism Breakdown
Qwen3.8-Max-Preview employs a sparse MoE architecture with a total of 2.4 trillion parameters, but the number of activated parameters has not been disclosed. Previously, Qwen3-235B-A22B activated 22 billion parameters, and Qwen3-30B-A3B activated approximately 300 million parameters, demonstrating that activated scale directly determines inference cost. Storing weights for 2.4 trillion parameters at 4-bit precision would require about 1.2TB. With a single Nvidia H200 having 141GB of VRAM, an eight-card configuration still faces practical deployment constraints.
The preview version is offered via API, compatible with OpenAI and Anthropic protocols, and inherits the 1M token context window from Qwen3.7-Max. Shuai Bai noted that the model is in a continuous state of evolution, and this is an early version.
Industry Impact
For developers, the 10% pricing lowers the barrier to API calls, but the lack of details on activated parameters and benchmarks makes it difficult to assess long-term self-hosting costs. Enterprise users can quickly test multimodal capabilities through the subscription, but self-hosting requires waiting for the open-weight version.
In terms of the competitive landscape, the simultaneous launch of Kimi K3 and Qwen3.8-Max-Preview shows that Chinese labs are advancing in parallel toward trillion-parameter models. The open-weight plan may accelerate controllable enterprise deployment and reduce reliance on closed-source models.
For upstream and downstream, compute demand will increase with total parameter scale, and pricing strategy needs to balance the preview discount with revenue from the official version.
Comparison and Precedents
Compared to Kimi K3, both models take an open-weight approach and have similar parameter scales. Qwen's history shows that MoE design can reduce activation costs, but this time the activation parameters were not disclosed, contrasting with previous versions.
Strategic Assessment
The most likely next step is for Alibaba to disclose the activated parameters or provide quantized checkpoints to lower the deployment threshold. Verification signals include the open-weight release timeline and the publication of actual benchmark data.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接