What it does
Qwen3 is a series of large language models from the Qwen team at Alibaba Cloud, released with public weights. The original release comes in dense sizes of 0.6B, 1.7B, 4B, 8B, 14B and 32B, plus mixture-of-experts models at 30B-A3B and 235B-A22B. Models switch between a thinking mode for reasoning, math and coding and a non-thinking mode for general chat. The Qwen3-2507 update adds separate Instruct and Thinking variants at 235B-A22B, 30B-A3B and 4B, with a 256K-token context that can be extended to 1 million tokens. The team also lists support for 100+ languages and dialects, and agent abilities such as tool use. The repository documentation covers running the models locally with Transformers, llama.cpp, Ollama, LM Studio and MLX, and serving them through an OpenAI-compatible API with SGLang, vLLM and TensorRT-LLM. It points to Qwen-Agent for tool use and to Axolotl, Unsloth, Swift and Llama-Factory for fine-tuning. All open-weight models are licensed under Apache 2.0.
Capabilities
- Dense and mixture-of-experts models from 0.6B to 235B-A22B
- Switchable thinking and non-thinking modes
- Instruct and Thinking 2507 variants with 256K context, extendable to 1M tokens
- 100+ languages and dialects
- Tool use through Qwen-Agent, with MCP support
- Apache 2.0 open weights
When to use it
- Running an open-weight model locally with llama.cpp, Ollama, LM Studio or MLX
- Self-hosting a model behind an OpenAI-compatible API with vLLM, SGLang or TensorRT-LLM
- Fine-tuning an open model on your own data
- Reasoning or long-context tasks where a thinking mode helps
