Qwen3-235B-A22B

Alibaba’s Qwen3-235B-A22B is a flagship Mixture-of-Experts language model in the Qwen3 family. It has 235B total parameters with 22B activated per token (A22B), 128 experts with 8 active, and ships under the Apache-2.0 license. The model’s hallmark is a built-in switch between “thinking” (emits a <think>…</think> trace for deeper reasoning) and non-thinking modes, plus strong agent/tool calling and multilingual coverage across 100+ languages.
For context length, the base A22B runs 32,768 tokens natively and ~131,072 with YaRN rope scaling. Alibaba also publishes tuned variants: Qwen3-235B-A22B-Instruct-2507 (non-thinking) with a 262K native window extendable to ~1.01M, and a Thinking-2507 line aimed at harder reasoning. Qwen3 models are supported in Hugging Face Transformers, vLLM, and SGLang, with guidance for local runtimes (Ollama, llama.cpp, MLX-LM).