Qwen3

Qwen3 is a series of open-weight large language models developed by the Qwen team at Alibaba Cloud, available in dense and Mixture-of-Experts (MoE) variants ranging from 0.6B to 235B parameters. The models support seamless switching between a thinking mode (for complex reasoning, math, and coding) and a non-thinking mode (for efficient general-purpose chat). Qwen3 supports 100+ languages and dialects and achieves state-of-the-art performance among open-weight models on reasoning, coding, and agent benchmarks. The latest Qwen3-2507 update extends long-context understanding to 256K tokens, extendable to 1 million tokens.

• Dense and MoE model sizes: Available in 0.6B, 1.7B, 4B, 8B, 14B, 32B (dense) and 30B-A3B, 235B-A22B (MoE) to fit a wide range of hardware budgets.
• Thinking and non-thinking modes: Switch between deep reasoning mode and fast chat mode using enable_thinking flags or /think//no_think instructions in the prompt.
• Long-context support: Handles up to 256K tokens natively, extendable to 1 million tokens with updated Qwen3-2507 model variants.
• Multilingual capability: Supports 100+ languages and dialects with strong multilingual instruction following and translation.
• Agent and tool use: Integrates with Qwen-Agent for tool use and MCP support, enabling precise function calling in both thinking and non-thinking modes.
• Broad inference framework support: Run with Transformers, vLLM, SGLang, TensorRT-LLM, llama.cpp, Ollama, LM Studio, MLX LM, OpenVINO, ExecuTorch, and MNN for flexible local and cloud deployment.
• Finetuning support: Compatible with Axolotl, UnSloth, Swift, and LLaMA-Factory for SFT, DPO, and GRPO training workflows.
• Quantization: Supports GPTQ, AWQ, and GGUF quantization for efficient deployment on consumer hardware.
• Apache 2.0 license: All open-weight models are freely available for commercial and research use.