BOSS Zhipin open-sources 33B Agent model: leads in 17 of 20 self-evaluations, but the ultimate goal is 3B.

Beating AI News Flash: Nanbeige Lab, under BOSS Zhipin, has released the open-source model Nanbeige4.3-33B, focused on coding and office Agents, supporting a 512K context window and licensed under Apache 2.0.

Officially, it is compared with Qwen3.8-27B, Meta Muse Glimmer 30B, and Gemma 4 31B. Among the 20 published tests, Nanbeige ranks first in 17. It scores 80.9 on Terminal-Bench 2.1, higher than Qwen's 73.0; DeepSWE reaches 65.5, while its rivals only hit 42.5. It also leads on the office automation test Automation Bench with 54.2 versus Qwen's 38.3. However, SWE-Bench Pro and GDPval are still led by Qwen.

Interestingly, Nanbeige previously focused on 3B-class small models, and the previous generation Nanbeige4.2-3B improved the model's effective computational depth without adding parameters by reusing Transformer layers. This time, the team instead first trained a 33B large model, intending to use it as a "teacher" and then distill capabilities such as coding and tool calling into the upcoming Nanbeige4.3-3B.