# Alibaba’s Qwen3 Aims to Disrupt the AI Landscape with ‘Hybrid’ Reasoning Models

## Alibaba’s Qwen3 Aims to Disrupt the AI Landscape with ‘Hybrid’ Reasoning Models

Alibaba has entered the ring with its latest offering in the AI space: Qwen3. Unveiled on Monday, April 28, 2025, Qwen3 is a family of large language models (LLMs) that the Chinese tech giant claims can rival, and in some instances surpass, the performance of leading models from Google and OpenAI.

What makes Qwen3 stand out? According to Alibaba, these are “hybrid” models, capable of both quick responses and in-depth reasoning for more complex tasks. This architecture allows the models to “think” through problems, improving accuracy and enabling self-fact-checking. The Qwen team highlights the seamless integration of thinking and non-thinking modes, giving users greater control over the computational resources allocated to each task.

The Qwen3 family spans a wide range of sizes, from a lean 0.6 billion parameters to a massive 235 billion parameters. As a general rule, a higher parameter count translates to improved problem-solving abilities in AI models. The majority of these models are either already available, or soon will be, under an open license on popular AI development platforms such as Hugging Face and GitHub. This commitment to open access distinguishes Alibaba’s approach.

This release comes at a pivotal time in the AI industry. The rise of powerful Chinese models like Qwen is placing increased pressure on American labs to innovate further and faster. It’s also spurred policymakers to enact regulations that restrict Chinese companies’ access to the advanced chips required for training these complex models, reflecting the escalating competition in the AI sector.

One of the defining characteristics of some Qwen3 models is their adoption of a Mixture of Experts (MoE) architecture. This approach enhances computational efficiency by breaking down complex tasks into smaller subtasks, delegating them to specialized “expert” models.

The Qwen3 models are multilingual, supporting 119 languages, and have been trained on a massive dataset of approximately 36 trillion tokens. This data encompasses a variety of sources, including textbooks, question-answer pairs, code snippets, and even AI-generated data.

Alibaba touts significant improvements in Qwen3’s capabilities compared to its predecessor, Qwen2. While not definitively surpassing the very latest models from OpenAI and others, the benchmarks suggest they are strong contenders.

For example, the largest Qwen3 model, Qwen-3-235B-A22B, reportedly edges out OpenAI’s o3-mini and Google’s Gemini 2.5 Pro on the Codeforces programming contest platform. It also outperforms o3-mini on the AIME math benchmark and the BFCL reasoning test. However, Qwen-3-235B-A22B is currently not publicly accessible.

The largest publicly available model, Qwen3-32B, remains highly competitive with several other open and closed AI models, including DeepSeek’s R1. Qwen3-32B surpasses OpenAI’s o1 model on various tests, including the LiveCodeBench coding benchmark.

Beyond raw performance, Alibaba emphasizes Qwen3’s proficiency in tool-calling, instruction following, and specific data format handling. Qwen3 is also available through cloud providers like Fireworks AI and Hyperbolic.

According to Tuhin Srivastava, CEO of AI cloud host Baseten, Qwen3 represents another step forward in the trend of open-source models catching up with proprietary systems like OpenAI’s offerings.

“The U.S. is doubling down on restricting sales of chips to China and purchases from China, but models like Qwen 3 that are state-of-the-art and open… will undoubtedly be used domestically,” Srivastava told TechCrunch. “It reflects the reality that businesses are both building their own tools [as well as] buying off the shelf via closed-model companies like Anthropic and OpenAI.”

With its combination of performance, open access, and hybrid reasoning capabilities, Qwen3 has the potential to significantly impact the future of AI development and deployment. It marks another step towards a more diverse and competitive AI landscape.