Nanbeige has introduced Nanbeige4.3-33B, a new open agentic coding model. The company plans to use this 33-billion parameter model as a teacher to distill agentic capabilities into smaller versions, beginning with the Nanbeige4.3-3B model.
The model utilizes the LlamaForCausalLM architecture without expert routing. It features specific controls for reasoning, including enable_thinking and preserve_thinking, alongside support for tool calling via an XML format. The model is compatible with Transformers, SGLang, and vLLM.
In internal benchmarks, Nanbeige compared its model against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. While the base model outperformed Gemma4-31B-Base across eight tasks, including MMLU-Pro and GPQA, the company noted that some competitor figures differ from official lab publications due to varying testing harnesses.
Technical requirements for the model include approximately 66 GB for BF16 weights. With a KV cache of about 208 KiB per token, a full 524,288-token sequence requires roughly 112 GB of memory. Currently, the model is not hosted by any Hugging Face inference provider.
