Qwen3 235B A22B

Qwen3-235B-A22B - Sparse Mixture-of-Experts Model

Qwen3-235B-A22B is the original mixture-of-experts (MoE) model in Qwen's Qwen3 series. Like its dense siblings, its documented identity centers on a built-in mechanism for moving between thinking mode and non-thinking mode inside one set of weights, rather than shipping these as separate models.

• Hard switch: the enable_thinking parameter, set in the chat template or API call, locks the model into thinking mode (which produces a visible <think>...</think> block before the answer) or non-thinking mode for the whole session.
• Soft switch: when thinking mode is enabled, a user can add /think or /no_think directly inside a message to change the mode turn by turn, and the model follows whichever tag appeared most recently in the conversation.

Sparse Routing Architecture

• Documented to route each token through 8 of 128 experts, the mixture-of-experts structure that distinguishes it from the dense Qwen3 models.
• Native context length of 32,768 tokens, validated up to 131,072 tokens using the YaRN scaling method.

Agent Use and Language Coverage

• Documented to work with Qwen-Agent for tool calling in both thinking and non-thinking modes, using MCP configuration files or built-in tools such as a code interpreter.
• Supports 100+ languages and dialects for instruction following and translation.