Nanbeige releases open-weight Nanbeige4.3-⁠33B under Apache 2.0, designed as a teacher for smaller agent models

The Nanbeige team released Nanbeige4.3-33B, an open-weight agent model, on Hugging Face under the Apache 2.0 license on October 25. The contact address on the model card belongs to Kanzhun, which operates the Chinese recruiting platform BOSS Zhipin. The team also published the base model, Nanbeige4.3-33B-Base. According to the model card, it is designed as the flagship teacher of the Nanbeige4.3 series. Its agent capabilities will be distilled into smaller models, starting with Nanbeige4.3-3B.

Per Hugging Face, the model has about 33.1 billion parameters and uses the Llama architecture, with a context length of 524,288 tokens. It runs with Transformers, SGLang and vLLM, and the card shows a setup split across four GPUs. The team compared it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. These are the team’s own benchmarks. Per the card, it scores 80.9 on Terminal-Bench 2.1, ahead of Qwen3.8-27B’s 73.0.

The model does not lead every test. It trails Qwen3.8-27B on SWE-Bench Pro, 60.3 to 61.7, and on GDPval, 84.8 to 86.4. On GPQA Diamond, Gemma4-31B scores 91.3 against its 89.5. Some tests were run with different agent harnesses such as Claude Code and mini-swe-agent, and one was graded with GPT-6 Sol as the judge. No independent evaluation has been published yet.