Ahead of DeepSeek’s R2 release, Alibaba announced its new Qwen3 AI model. The Chinese giant wants to improve its artificial intelligence capabilities and efficiency compared to the DeepSeek model.
According to Nikkei Asia, the e-commerce company reduced computing power while still being able to train 36 trillion tokens, double the amount used for its previous Qwen2.5 model. With that, Alibaba’s updated model can switch between a reasoning model for complex tasks and a faster mode for quick, everyday queries.
The latest open-source LLM is Alibaba’s first set of hybrid reasoning models. By using dense models for creating tasks and two Mixture-of-Experts (MoE) models, Alibaba is able to build its model for much less than the cost of Google’s Gemini, OpenAI’s ChatGPT, and Microsoft’s Copilot.
As usual for new releases, Alibaba claims its top Qwen3 model can outperform models released between December and March, including DeepSeek-R1, OpenAI-o1, Grok-3, and Gemini-2.5-Pro in benchmark tests for mathematical reasoning, coding proficiency, and tool nd function calling capabilities.
The publication says it took Alibaba seven months of work to release this AI model. Interesting, this release comes a few days after Baidu released two new AI models that cost “a fraction” of DeepSeek options.
What makes these releases so interesting is that Chinese companies are making their AI models cheaper while exponentially improving their reasoning capabilities. While this could help make the technology more affordable, it’s also necessary for these companies as the US “tightens restrictions on chip exports to curb China’s advances in the field.”
With cheaper and smarter AI models, OpenAI, Perplexity, Google, and many others also need to improve their LLMs to make them more powerful without costing users more. When DeepSeek released its R1 model, it stunned the world and crashed the US stock market in the process.
After that, the American companies started to speed up the development of their latest models while still making the latest technologies available for more users.
Soon, we’ll see if the DeepSeek R2 model will change everything one more time, or if OpenAI and other companies still have some advantage before they release their next breakthrough models.
