On Tuesday, Alibaba released the third generation of its Qwen family of AI models, with multiple versions, each featuring a different number of parameters. The largest model, with 235 billion parameters, outperformed DeepSeek-R1 and OpenAI’s o1 reasoning models, according to Alibaba, owner of the Post. At 600 million parameters, the most efficient version of the model could be capable of running on a smartphone, some experts have said.
Hours after Alibaba released Qwen3 to the open-source AI development platform Hugging Face, Tesla CEO Musk posted to his microblogging site X that his start-up xAI would release an early beta version of Grok 3.5 to SuperGrok subscribers, who get premium access to the Grok chatbot. “It is the first AI that can, for example, accurately answer technical questions about rocket engines or electrochemistry,” Musk wrote.
The launch of the DeepSeek-R1 model in January marked the beginning of a renewed AI race that has seen an accelerated release schedule of new models, with many focusing on energy efficiency. The low-cost, high-performing DeepSeek models were seen as a wake-up call to US developers, suggesting the country’s AI industry was not as far ahead as it once thought.
In addition to Alibaba, internet search giant Baidu, ByteDance and Tencent Holdings have updated their foundational models in the last three months, nearing or matching the performance of American models such as Google’s Gemini 2.5 Pro, OpenAI’s o3 and o4 and Meta Platforms’ Llama 4.
A report by Stanford University this month found that China has rapidly narrowed the gap with the US in producing cutting-edge AI models. China’s open-source models have also proven popular with developers and users. Alibaba’s Qwen is currently the world’s largest open-source AI ecosystem, with over 100,000 derivative models, exceeding those based on Meta’s Llama.
