Alibaba Group has launched its latest artificial intelligence (AI) breakthrough — the Qwen3 series — positioning it as a serious contender in the ever-intensifying global AI arms race. The third generation of the Chinese tech giant’s open-source large language models (LLMs) was unveiled on Tuesday and claims to outperform major players such as DeepSeek-R1, OpenAI’s o1, and Google’s Gemini.
According to Alibaba’s cloud division, the Qwen3 series brings substantial improvements in processing speed, multilingual understanding, and reasoning capability, with the company asserting that it marks a pivotal moment in its pursuit of artificial general intelligence (AGI) and eventually, artificial superintelligence.
What is the Qwen3 series?
The Qwen3 family comprises eight models, ranging in size from 600 million to a colossal 235 billion parameters. These parameters are key to a model’s intelligence, dictating how well it performs complex tasks such as natural language processing, computer programming, and mathematical problem-solving.
Alibaba says each of these models is optimised for specific use cases, from lightweight mobile deployment to heavy-duty research and enterprise integration. Particularly noteworthy is the Qwen3-235B, the flagship of the series, which reportedly demonstrates world-leading performance.
Qwen3 vs the competition
Benchmarking results published by Alibaba suggest that Qwen3-235B and the compact Qwen3-4B model match or even exceed the performance of DeepSeek-R1, OpenAI’s o1, and Google Gemini. Testing included tasks such as instruction following, creative text generation, coding support, mathematical analysis, and multi-step logical reasoning.
In a blog post, the Qwen team remarked:
“Qwen3 represents a significant milestone in our journey towards artificial general intelligence and artificial superintelligence. Through enhanced pre-training and advanced reinforcement learning methods, we’ve achieved a new level of capability and efficiency.”
The team also highlighted that their Mixture-of-Experts (MoE) model, the Qwen3-30B-A3B, even outperforms the previous QwQ-32B, while the smaller Qwen3-4B gives the older and significantly larger Qwen2.5-72B-Instruct a run for its money — an achievement that underlines the series’ improved efficiency.
The hybrid reasoning edge
A standout feature of the Qwen3 models is their “hybrid reasoning” system. Unlike many rival AIs that use a single problem-solving approach, Qwen3 allows users to toggle between two modes:
- Thinking mode: Slower but more comprehensive, ideal for nuanced or high-stakes tasks
- Non-thinking mode: Faster, for straightforward questions and casual interactions
By comparison, DeepSeek-R1 utilises a more rigid Chain-of-Thought (CoT) reasoning approach, where the model generates step-by-step logic before delivering an answer. Qwen3’s hybrid method offers greater flexibility and customisation.
Multilingual capabilities and training scale
Alibaba also revealed that Qwen3 models were trained on a massive 36 trillion tokens drawn from texts in 119 languages and dialects — a threefold increase over Qwen2.5. This multilingual boost is expected to significantly improve the model’s understanding of culturally diverse and region-specific queries, giving it an edge in global applications.
More Read
How to access and use Qwen3
The Qwen3 models are now accessible via popular AI hosting platforms such as Hugging Face, ModelScope, Kaggle, and GitHub (via Microsoft). Developers and researchers can deploy Qwen3 using frameworks like SGLang and vLLM, while tools like Ollama, LMStudio, MLX, llama.cpp, and KTransformers support local integration.
Global AI competition heats up
The Qwen3 launch is Alibaba’s boldest move yet in the fiercely competitive AI landscape, which is seeing rapid innovation from Chinese giants like Baidu and Tencent, as well as international leaders such as Anthropic, Google DeepMind, and Meta AI. Baidu, notably, has just upgraded its own ERNIE models, while DeepSeek is expected to reveal DeepSeek-R2 shortly.
With Qwen3, Alibaba hopes to reassert its dominance in AI and offer a compelling, open-source alternative to the closed systems developed in Silicon Valley.
Whether the new series lives up to its lofty claims remains to be seen — but one thing is clear: the global AI race has just accelerated.
