BOSS Zhipin open-sourced the 33B Agent model, leading in 20 self-evaluations and 17 items

The recruitment platform BOSS Zhipin's Nanbeige Laboratory has released the open-source model Nanbeige4.3-33B, aimed at programming and office agents, supporting 512K context, and using the Apache 2.0 license. The official comparison with Qwen3.8-27B, Meta Muse Glimmer 30B, and Gemma 4 31B shows that it ranked first in 17 out of 20 tests published.

Terminal-Bench 2.1 scored 80.9, higher than Qwen's 73.0; DeepSWE reached 65.5, with a control score of 42.5; the office automation test Automation Bench scored 54.2, higher than Qwen's 38.3. SWE-Bench Pro and GDPval are still led by Qwen. Previously, Nanbeige mainly developed 3B level small models, and the previous generation Nanbeige4.2-3B improved effective computational depth without increasing parameters by reusing Transformer layers. This time, the 33B model was trained first, preparing it to act as a teacher, and then distilling capabilities such as programming and tool invocation to the upcoming Nanbeige4.3-3B.