Alibaba Previews 2.4-Trillion-Parameter Qwen3.8-Max Model
Alibaba’s Qwen team is expanding its push into generative artificial intelligence as Chinese developers compete with domestic and global rivals on increasingly capable foundation models. Qwen3.8-Max is designed to process text, images, video and documents, reflecting an industry shift beyond standalone language systems toward multimodal tools suited to research, content creation and enterprise workflows.
Alibaba previewed Qwen3.8-Max-Preview at the World Artificial Intelligence Conference in July 2026, presenting it as a 2.4-trillion-parameter multimodal model and claiming performance behind only Fable 5. The unveiling came just days after Moonshot AI launched its Kimi K3 open-weight model, underscoring the accelerating release cycle among Chinese AI companies seeking an edge with larger flagship systems.
All Coverage
3 original reportsThe Backstory
The history behind this eventAlibaba Unveils Qwen3.8 Model as Preview of Qwen4 Architecture
Alibaba’s Qwen team has been expanding its open-source artificial intelligence ecosystem while pursuing models that lower inference costs without sacrificing capability. Qwen3.8-Flash-Next is positioned as a technical preview of the architecture planned for the next-generation Qwen4 flagship, with Alibaba targeting practical workloads including software coding and office productivity where speed and operating costs are key considerations.
The Qwen team has released Qwen3.8-Flash-Next, a multimodal mixture-of-experts model with 125 billion total parameters but only 6 billion activated for each inference task. The sparse design is intended to reduce computing requirements while retaining strong performance. The release gives developers an early look at Qwen4’s architectural direction, though the reports provided no launch date or detailed benchmark results for the eventual flagship model.
Alibaba Open-Sources Qwen3.8-27B, Touts Visual Edge Over Claude
Alibaba Group’s Qwen3.8-27B is a 27-billion-parameter, natively multimodal model designed to process both text and visual inputs. Its release under the Apache 2.0 license lowers barriers for developers and companies seeking to modify, deploy and commercialize the technology. Quantized versions can run on consumer-grade graphics cards, broadening local access to capable visual AI and reducing reliance on costly cloud infrastructure as open models compete with proprietary systems.
As of Aug. 16, 2026, benchmarks published by Alibaba showed Qwen3.8-27B outperforming Anthropic’s Claude Opus 4.6 Max on nine measures, including visual-agent tasks and chart understanding. The results point to strong capabilities in interpreting interfaces and acting on visual information. However, the model remained behind Claude in pure-text reasoning and long-chain coding, suggesting its advantage is concentrated in visual operation while deeper logic and complex code generation remain areas for improvement.
Alibaba Launches Qwen3.8-Max, Plans Open-Weight Release
Alibaba has expanded its Qwen family as Chinese technology companies intensify efforts to narrow the gap with leading U.S. artificial-intelligence developers. Qwen3.8-Max uses a mixture-of-experts, or MoE, architecture with 2.4 trillion total parameters and is designed to improve coding and office-productivity tasks. Its scale, Alibaba Cloud distribution and planned open-weight release could broaden access for developers while strengthening Alibaba’s position in enterprise AI.
Alibaba launched Qwen3.8-Max on Aug. 5, 2026, saying benchmark results put its performance on par with models from Anthropic, the maker of Claude. Developers can already access the model through an API on Alibaba Cloud. The company plans to publish the model weights on Hugging Face and ModelScope during the week of Aug. 10, extending its strategy of combining commercial cloud services with publicly available AI models.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.
If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →