Nanbeige has released Nanbeige4.3-33B for coding and professional agent workflows. Its model card documents up to 512K context, configurable retention of prior reasoning, XML and JSON tool-call formats, and deployment examples for vLLM and SGLang. It recommends retaining reasoning for multistep tool use and positions the model as a teacher for future smaller distilled models. Published benchmarks use the team’s specified reasoning settings and execution harnesses; they are not independent proof of superiority on every coding task.