Qwen 3 235B Instruct

Qwen 3 235B A22B Instruct (2507) is Alibaba's large Mixture-of-Experts model, activating about 22B of 235B parameters per token for efficient inference at scale. This non-thinking variant prioritizes direct responses and is tuned for agentic workflows, tool use, multilingual tasks, and complex question answering across academic, coding, and reasoning benchmarks. It reports strong results including 70.3 on AIME 2025 mathematics and 87.9 on MultiPL-E coding, with a 128K context window. Released under Apache 2.0, it is well suited to long-context analysis, multi-step reasoning, and production deployments where the MoE design keeps inference cost manageable relative to dense models of similar capability.