Alibaba Unveils Its Most Capable AI Model Yet, Trailing Just Behind Moonshot in Size
Alibaba unveiled Qwen3.8-Max on Monday, its largest AI model yet, weeks after rival Moonshot's Kimi K3 rattled markets and pressured Alibaba to prove its latest flagship could finally back up earlier claims with verified results
Alibaba’s timing for the Qwen3.8-Max release wasn’t accidental.
Chinese AI labs have spent the summer locked in an increasingly public contest to prove they can match Western frontier models without the vast computing budgets, and Monday’s launch gave Alibaba a chance to reclaim ground it had ceded to Moonshot’s Kimi K3 last month.
Unlike an earlier preview Alibaba floated in July, this release arrived with independently checkable leaderboard placements attached, a detail that mattered given how the last announcement played out.
A Close Race on Size, a Real Test on Substance
Reuters reported that Qwen3.8-Max’s 2.4 trillion parameters sit not far behind Moonshot’s 2.8 trillion, a gap that matters less for raw capability than the scale of computing and data behind each AI system.
Alibaba said Qwen3.8 Max uses a mixture-of-experts design, similar to Google’s DiffusionGemma, activating only 95 billion parameters per task to improve efficiency and reduce costs.
Reuters said both Qwen3.8 Max and Kimi K3 can process text, images, and video, with one million token context windows that let them analyse lengthy documents, codebases, or legal files in a single pass.
The Claim That Wasn’t Verified the First Time
At July’s World Artificial Intelligence Conference Shanghai, framed as a showcase for China’s AI leadership, Alibaba claimed Qwen3.8-Max trailed only Fable 5 among world-leading models.
However, the company did not publish benchmark scores, a model card, or independent test results to support that claim. But subsequent Arena.AI results revealed a notable gap from those early promises.
On the text leaderboard, the fully launched Qwen3.8-Max sits behind Claude Fable 5 and three separate Opus variants, though on the vision benchmark it holds closer to its original hype, trailing directly below Fable 5.
A 16-Day Coding Marathon and an Open-Weight Bet
Bloomberg, via Yahoo Finance, highlighted a feature Alibaba is using to set Qwen3.8 Max apart beyond benchmark scores.
During internal testing, the model reportedly completed a software engineering project on its own over 16 days, demonstrating long-term autonomous coding that Alibaba sees as a key advantage.
Bloomberg also confirmed Alibaba will publicly release Qwen3.8 Max’s weights next week, keeping it alongside Moonshot, DeepSeek, and ByteDance in the open weight camp, even as DeepSeek expanded access to its V4 Flash model last week.
That crowded release schedule highlights how quickly China’s AI sector is advancing. Despite ongoing distillation allegations, each new model is now judged not only against Western frontier labs but also against the latest release from domestic rivals.
Source: Alibaba Releases AI Model With Performance Rivaling Anthropic
