How does a 33B model compare with other open models around the same size on agentic coding?
Nanbeige has open-sourced Nanbeige4.3-33B under Apache 2.0 with a 512K context window—the amount of text a model can take into account at once.
According to its model card, Nanbeige claims it leads open ~30B models on agentic coding, where models take actions to carry out coding tasks. It reports 80.9 on Terminal-Bench 2.1, compared with 73.0 for Qwen3.8-27B. These are vendor-reported results, so the comparison should be read with that attribution attached.
Nanbeige is also positioning the model as a teacher for smaller models, starting with a 3B model. The intended direction is to pass capabilities from a larger model to a smaller one; the announcement does not establish how much capability that smaller model will retain.
For practitioners, the useful next step is to evaluate the released model on representative coding tasks. Treat the reported benchmark score and 512K context as separate things to test: a high score does not establish performance on a particular workflow, and context capacity alone does not establish how effectively the model uses that information.