Alibaba Cloudâs Qwen3: The Rivalling Open-Source AI Model
As Western companies like OpenAI, Anthropic and Google DeepMind continue to advance their AI proprietary models, China's technology giants have pursued parallel development paths with a focus on open-source alternatives that can serve both domestic and international markets.
In this context, Alibaba Cloud has launched Qwen3, the latest generation of its open-sourced large language model (LLM) family that introduces hybrid reasoning capabilities to the marketplace.
The series features six dense models and two Mixture-of-Experts models, providing developers with options to build applications across various hardware platforms from mobile devices to autonomous vehicles.
Dense models, which use all parameters for every input, range from 0.6B to 32B parameters, while the MoE models – systems that selectively activate only a subset of parameters for each input – include a 30B model with 3B active parameters and a 235B model with 22B active parameters.
All models in the series are now open-sourced and available globally through popular AI development platforms.
Alibaba’s Qwen3 Models introduce Thinking Mode
Qwen3 is Alibaba's introduction of hybrid reasoning models that can alternate between two operational states: a thinking mode for complex, multi-step tasks such as mathematics and coding and a non-thinking mode for faster, general-purpose responses.
Thinking mode allows the model to perform extended reasoning with context lengths of up to 38,000 tokens, enabling developers to control the trade-off between performance and computational efficiency.
Meanwhile, “the Qwen3-235B-A22B MoE model significantly lowers deployment costs compared to other state-of-the-art models, reinforcing Alibaba’s commitment to accessible, high-performance AI,” Alibaba says, aligning with its strategy to make advanced AI technologies more accessible to developers worldwide.
- Thinking Mode: In this mode, the model takes time to reason step by step before delivering the final answer
- Non-Thinking Mode: Here, the model provides quick, near-instant responses, suitable for simpler questions where speed is more important than depth
Trained on a dataset of 36 trillion tokens – twice the size used for its predecessor Qwen2.5 – the new models demonstrate improvements in reasoning, instruction following, tool integration and multilingual capabilities.
Qwen3 supports 119 languages and dialects, positioning it as a versatile option for global applications requiring translation and multilingual functionality.
Qwen3’s performance benchmarks
The models achieve competitive results across industry benchmarks including AIME25 for mathematical reasoning, LiveCodeBench for coding abilities, BFCL for tool and function-calling capabilities and Arena-Hard for instruction-tuned language models.
Alibaba implemented a four-stage training process to develop the hybrid reasoning model, incorporating long chain-of-thought cold start, reasoning-based reinforcement learning, thinking mode fusion and general reinforcement learning.
Qwen3 models are now available for download on Hugging Face, Github and ModelScope, with API access forthcoming through Alibaba's AI model development platform, Model Studio.
The models also power Alibaba's AI assistant application, Quark.
Since its introduction, the Qwen model family has recorded over 300 million downloads worldwide, with developers creating more than 100,000 Qwen-based derivative models on Hugging Face, establishing it as one of the most widely adopted open-source AI model series globally.
Explore the latest edition of AI Magazine and be part of the conversation at our global conference series, Tech & AI LIVE.
Discover all our upcoming events and secure your tickets today.
Also sign up to our free weekly newsletter for the latest insights and stories straight into your inbox.
AI Magazine is a BizClik brand
