Nanbeige Releases 33B Parameter Model Under Apache 2.0 License
The 33-billion-parameter Nanbeige4.3-33B model has been released under an Apache 2.0 license featuring a 512K-token context window and performance metrics for agent-based tasks.
Nanbeige4.3-33B is a 33-billion-parameter model released under the Apache 2.0 license. You can download the weights from Hugging Face. The team built it as a teacher for planned smaller agent models, which include an upcoming 3B version with no release date yet.
The model card lists a 512K-token context window. Because of the Apache 2.0 license, you can modify the weights and use them commercially.
Nanbeige reports 80.9 on Terminal-Bench 2.1, while Qwen3.8-27B scores 73.0. The team states they ran both models using the same harness and inference settings. However, the results are mixed. On SWE-Bench Pro, Nanbeige4.3-33B scores 60.3 against 61.7 for the Qwen model. These figures are vendor-reported numbers from the releasing team's own evaluations and have not been replicated by any independent group.
Reproducing the agent results requires a specific setting. Evaluations ran with preserve_thinking=true, which retains earlier reasoning in the context across turns. The model card recommends this for multi-turn tool use, though not for general chat. A fairer comparison would hold that setting fixed and measure both task success and total generated tokens, since retained reasoning can increase token usage.