Nanbeige4.3-33B: an open 33B-parameter model for agentic tasks

The model Nanbeige/Nanbeige4.3-33B with the text-generation task has been published on HuggingFace. The author positions it as the flagship "teacher" model in the ~30B parameter class: it is planned to be used to train and distill lighter agentic models, including Nanbeige4.3-3B. The model card provides comparisons with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on the Terminal-Bench, DeepSWE, SWE-Bench Pro and other benchmarks.

What it means

The model contains 33.1 billion parameters; the full set of weights in the repository takes up 61.7 GB in safetensors format. License — apache-2.0. It runs via transformers. 49 quantized builds have already been posted on the platform: GGUF, NVFP4, FP8, MLX, EXL, AWQ, GPTQ. The most compact working build is IQ4_XS at 16.8 GB, meaning the model can realistically run on a single consumer GPU with 24 GB of VRAM or on a CPU with enough system RAM. Running the full model locally in FP16 will require noticeably more resources than the quantized variants. The author does not specify minimum VRAM requirements and provides no details about the context window. Judging by the fact sheet, the practical scenario is to take a GGUF build at the IQ4_XS level or similar in size. The model is of interest primarily to those working with agentic scenarios and complex coding, rather than simply generating text. More news on local AI tools is in the CPU3D news feed.