Alibaba launches Qwen3, the latest generation of its open-source large language mode (LLM) family, setting a new benchmark for AI innovation.
The Qwen3 series includes six dense models (0.6B, 1.7B, 4B, 8B, 14B, and 32B parameters) and two Mixture-of-Experts (MoE) models (30B with 3B active, and 235B with 22B active), all now available globally for developers to build applications across mobile, wearables, autonomous vehicles, and robotics.
Hybrid Reasoning Breakthrough
Qwen3 introduces hybrid reasoning, combining traditional LLM capabilities with advanced dynamic reasoning. Models switch between thinking mode for complex tasks and non-thinking mode for fast, general-purpose responses.
API users can control thinking duration (up to 38K tokens), optimizing between performance and efficiency. The Qwen3-235B-A22B MoE model significantly reduces deployment costs compared to other state-of-the-art models.
Key Advancements
Trained on 36 trillion tokens (double its predecessor), Qwen3 delivers significant improvements in:
- Multilingual Mastery: Supports 119 languages and dialects with leading translation performance
- Advanced Agent Integration: Native Model Context Protocol (MCP) support and robust function-calling
- Enhanced Reasoning: Surpasses previous Qwen models (QwQ in thinking mode and Qwen2.5 in non-thinking mode) in mathematics, coding, and logic
- Enhanced Human Alignment: More natural creative writing, role-playing, and dialogue experiences
These advancements result from improved architecture, expanded training data, and more effective training methods, achieving top results across benchmarks including AIME25 (mathematical reasoning), LiveCodeBench (coding proficiency), BFCL(tool and function-calling capabilities), and Arena-Hard (benchmark for instruction-tuned LLMs).
Additionally, to develop the hybrid reasoning model, a four-stage training process was implemented, which includes long chain-of-thought (CoT) cold start, reasoning-base reinforcement learning (RL), thinking mode fusion, and general RL.
Open Access Ecosystem
Qwen3 models are available on Hugging Face, Github, and ModelScope, with API access coming soon through Model Studio. The models also power Alibaba’s AI assistant, Quark.
With over 300 million downloads and 100,000+ derivative models, Qwen has become one of the world’s most widely adopted open-source AI model series.
