The new Qwen3 family of ALIBABA AI models exceeded Deepseek R1 to become the best open source model in the world. According to reports, Qwen3 has done better than R1 in tests that measure the capacities of open source models in fields such as language teaching, mathematics, coding and data analysis.
The Qwen3 family was launched last week by Alibaba's cloud computing unit. It has eight improved models with between 600 and 235 billion parameters. In automatic learning, the parameters are the variables in an AI system during its training.
According to Lively Platform, an independent platform that tests large language models, before these new tests, the DEEPSEEK R1 had been the best model of Open-Source IA in the world since its release in January. But no more now.
American and Chinese companies rush to adopt Qwen 3
The rise of Qwen3 in the LiveBench classification shows how speed is developing in China. The Chinese technology industry has grown up a lot thanks to open-source tools. The Open-Source Alibaba method code has enabled other third-party software developers to share design, repair broken links or make the program more powerful.
However, the global results of Livebench showed that Qwen3 was not as good as O3 of Openai, Google's Gemini Pro 2.5 and Claude 3.7 of Anthropic, which are the best models of AI at the world. Livebench says that the more popular IA model of O3-Mini, the most popular IA model in Openai, was the best in the world as a whole. Microsoft Openai backup.
For each million tokens, it takes $ 10 to operate O3. On the other hand, Qwen3 is cheaper to use because it costs only $ 0.55 per 1 million tokens to execute. Because Qwen3 is cheaper and works better, many companies have declared that they would support the new Alibaba AI model as soon as it is released.
Huawei Technologies, Moore Threads, Cambconne Technologies and Hygon Information Technology are all flea companies that have said that they will support Qwen3.
Cambricon said last Tuesday that he managed to optimize QWEN3 to quickly operate on his graphic processing units. This was done because the developers of AI in the Philippines wanted fleas made in China.
Qwen3 is also used on the Hyperbolic and Fireworks .i -IA -infrastructure cloud computing services. Nvidia and Intel manufacturers have started supporting Qwen3.
Many Big Data centers in China, such as those of Beijing, Shanghai, Hangzhou and the provinces of Hubei, Jilin and North West of Shaanxi, also said that they would use the Alibaba third generation QWEN models. The SuperCalculculculé network in China also adopted Qwen3. This network connects more than 20 data centers in 20 cities in 14 provinces.
The CEO of Anthropic says that Deepseek was “a little exaggerated”
During a commercial event, a co-founder of Anthropic, the company that made the Claude AI models, said that Deepseek was still “six to eight months behind on American border companies”. He also said that the recent buzz around the Chinese start-up was “perhaps a little exaggerated”.
In depth Attracted attention around the world at the end of December 2024 and early January 2025 by sharing two models of advanced open source, V3 and R1. These models have been designed for a small fraction of the cost and computing power whose large technology companies generally need LLM projects.
It is not known when Deepseek will publish the next generation of his models. The company based in Hangzhou has quietly released its prover-V2 parameter of $ 671 billion April. It was an update of his specialized model to manage mathematical evidence. However, he said nothing about the progress of his long -awaited R2 reasoning model.
Cryptopolitan Academy: Do you want to develop your money in 2025? Learn to do it with DEFI in our next webclass. Record your place
