Nanbeige releases Nanbeige4.3-33B open-weight teacher model
Nanbeige released Nanbeige4.3-33B under the Apache 2.0 license, positioning it as a teacher model for smaller agents, including a planned Nanbeige4.3-3B. The model card lists a 512K context window, while Nanbeige reported 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B under the same settings. The results were not independently replicated, and the model scored 60.3 versus Qwen3.8-27B’s 61.7 on SWE-Bench Pro.
The release adds a permissively licensed, roughly 30B-parameter teacher model to the open-weight ecosystem and establishes the basis for Nanbeige’s planned smaller agent models, although its benchmark leadership remains based on vendor-reported evaluations.
Nanbeige released Nanbeige4.3-33B as the teacher for its planned smaller agents, including an upcoming 3B model. the weights are Apache 2.0, and the model card lists a 512K context window.
the team reports 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B, using the same harness and inference settings within each benchmark. it does not win everywhere. on SWE-Bench Pro, the reported scores are 60.3 versus 61.7. these are the releasing team's evaluations, not independent replication.
one practical detail for reproducing the agent results: the evaluations retain earlier reasoning with preserve_thinking=true. the card recommends that setting for multi-turn tool use, but not general chat. comparing task success and total generated tokens with that setting held fixed would be more useful than treating the headline scores as a deployment ranking.
#Nanbeige#Nanbeige4.3-33B#Nanbeige4.3-3B#Qwen3.8-27B#Muse-Glimmer-30B#Gemma4-31B#Terminal-Bench 2.1#SWE-Bench Pro