Qwen3 is a series of large language models from Alibaba Cloud, offering dense and Mixture-of-Experts (MoE) architectures with enhanced reasoning, human alignment, and multilingual capabilities. It targets developers and researchers seeking advanced LLM performance for chat, coding, and agentic tasks.

How It Works

Qwen3 models feature a novel "thinking mode" that can be seamlessly switched with a general-purpose "non-thinking mode" via chat templates or specific instructions. This dual-mode approach optimizes performance for complex reasoning tasks (math, code) and efficient general chat, surpassing previous Qwen versions and other open-source models in benchmarks.

Quick Start & Requirements

• Transformers: pip install transformers>=4.51.0
• llama.cpp: llama.cpp>=b5092
• Ollama: ollama serve (v0.6.6+)
• vLLM: vllm>=0.8.5
• SGLang: sglang>=0.4.6.post1
• MLX LM: mlx-lm>=0.24.0 (for Apple Silicon)
• Documentation: EN|ZH
• Models: Hugging Face or ModelScope

Highlighted Details

• Supports dense and MoE models across various sizes (0.6B to 235B).
• Excels in reasoning, human preference alignment, and multilingual tasks (100+ languages).
• Strong agent capabilities with precise tool integration.
• Offers OpenAI-compatible APIs via vLLM, SGLang, and Ollama.

Maintenance & Community

• Active development by Alibaba Cloud's Qwen team.
• Community channels: Discord, WeChat.

Licensing & Compatibility

• Licensed under Apache 2.0, permitting commercial use and closed-source linking.

Limitations & Caveats

• llama.cpp's chat template engine may encounter errors with Qwen3's specific template features; a fix is in progress.
• Ollama's default context management settings may require adjustment for optimal Qwen3 performance.