A man walks past the Alibaba logo displayed at a booth at the China International Supply Chain Expo in Beijing, China, on July 16, 2025.
Florence Law | Reuters
meta We want to lead in open source AI. chinese alibaba It just reminded American tech giants just how competitive the market is.
On Monday, the Chinese tech giant announced AI models designed to run on consumer hardware such as laptops, a new battleground for AI model developers, and published the weights of its most powerful models as it aims to maintain its lead in open source technology.
This comes after Meta last week announced plans to open source its most powerful AI model and launch a new model designed to run on laptops, aiming to position itself as a U.S. alternative to Chinese technology against Frontier Labs’ OpenAI and Anthropic, which take more proprietary approaches.
At the launch of the new model, Alibaba said the Qwen3.8-27B has “excellent capabilities” in coding, professional work, research, and handling long-term agent tasks. It rivals the performance of another model 10 times its size, it added.
Alibaba also said it has released the weightings (the calculations and rules that determine how the AI functions and behaves) for its most powerful model, Qwen3.8 Max. This means that you can download and run it freely. However, the data and methods used to train that model may not be revealed.
Alibaba has established itself as a leader in indiscriminate AI, and other Chinese companies such as DeepSeek and Moonshot are also strong players.
Meta was an early player in open source AI with the Llama family of models, but was quickly overtaken by Chinese players. Last week’s announcement included an open source model family called Muse Glimmer that is designed to run on laptops.
Nick Patience, head of AI at Futurum Group, told CNBC: “Meta’s re-adoption of promiscuous weights was itself a response to Chinese research institutions gaining significant share of the promiscuous weight market over the past two years.”
Alibaba is looking to strengthen its dominance
The popularity and success of an openweight model is often determined by the number of times the model is downloaded and the number of developers using it to build their own products.
Hugging Face, one of the largest repositories where developers can download openweight models, announced last week that Qwen-based models now account for 151,448 derived models (the number of times openweight AI models are downloaded and used). According to Huggingface, this is 2.6 times the total footprint of Meta.
Neil Shah, co-founder of Counterpoint Research, told CNBC that “the companies that can offer the most capable open weight models will advance in this race.”
“Alibaba aims to surpass meta and become the undisputed leader in this, a strong alternative to Silicon Valley’s frontier-grade deployable models, and has its sights set on global markets.”
The release of Qwen3.8-27B suggests that Alibaba expects more advanced models to run at the “edge” rather than pure data centers. Industry experts say on-device AI can run faster and be more secure because it runs on local hardware, such as a phone or laptop.
Counterpoint Research’s Shah said the “next battleground” for AI models is successfully running them on devices rather than from data centers. Futurum Group’s Payens said Alibaba has advantages in several areas, from open-ended models to on-device AI.
“Alibaba has made Qwen the most trusted non-US model family for building hardware relationships with the promiscuous developer community in China and around the world,” Patience told CNBC.
