Alibaba’s Qwen3 unseats DeepSeek’s R1, now leads open-source AI models

Alibaba’s new Qwen3 family of AI models has surpassed DeepSeek’s R1 to become the world’s best open-source model. According to reports, Qwen3 did better than R1 in tests that measure open-source AI models’ abilities in areas like language instruction, math, 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 million and 235 billion parameters. In machine learning, parameters are the variables in an AI system while it is being trained.

According to LiveBench platform, an independent platform that tests large language models, before these new tests, DeepSeek’s R1 had been the best open-source AI model in the world since it came out in January. But not anymore.

Both US and Chinese companies rush to adopt Qwen 3

The rise of Qwen3 in the LiveBench rankings shows how quickly AI is developing in China. The Chinese tech industry has grown a lot thanks to open-source tools. The Alibaba open-source method code has allowed other third-party software developers to share the design, fix broken links, or make the program more powerful. 

However, the overall LiveBench results showed that Qwen3 was not as good as OpenAI’s o3, Google’s Gemini Pro 2.5, and Anthropic’s Claude 3.7, which are the best closed-source AI models in the world. LiveBench says that the o3-mini high, OpenAI’s most popular AI model, was the best in the world overall. Microsoft backs OpenAI.

For every 1 million tokens, it takes $10 to run o3. On the other hand, Qwen3 is cheaper to use because it only costs $0.55 per 1 million tokens to run. Because Qwen3 is cheaper and works better, many businesses said they would back Alibaba’s newest AI model as soon as it came out.

Huawei Technologies, Moore Threads, Cambricon Technologies, and Hygon Information Technology are all chip companies that have said they will support Qwen3.

Cambricon said last Tuesday that it had successfully optimized Qwen3 to run quickly on its graphics processing units. This was done because AI developers in the Philippines wanted chips made in China.

Qwen3 is also being used on the cloud computing services of Hyperbolic and Fireworks.ai, two AI infrastructure companies. The American chipmakers Nvidia and Intel have begun to support Qwen3.

Many big data centers in China, like those in Beijing, Shanghai, Hangzhou, and the provinces of Hubei, Jilin, and Northwest Shaanxi, have also said they will use Alibaba’s third-generation Qwen AI models. The Supercomputing Network in China has also adopted Qwen3. This network links over 20 data centers in 20 towns across 14 provinces.

Anthropic CEO says that DeepSeek was “a bit overblown”

At a business event, a co-founder of Anthropic, the company that made the Claude AI models, said that DeepSeek is still “six to eight months behind where US frontier companies are.” He also said that the recent buzz around the Chinese start-up was “perhaps a bit overblown.”

DeepSeek got attention around the world in late December 2024 and early January 2025 by sharing two advanced open-source AI models, V3 and R1. These models were made for a small fraction of the cost and computing power that big tech companies usually need for LLM projects.

It’s unclear when DeepSeek will release the next generation of its models. The Hangzhou-based company quietly released its 671-billion-parameter Prover-V2 in late April. This was an update to its specialized model for handling math proofs. However, it hasn’t said anything about the progress of its long-awaited R2 reasoning model.

Cryptopolitan Academy: Coming Soon - A New Way to Earn Passive Income with DeFi in 2025. Learn More