Nanbeige launches 33b-parameter open agentic coding model

Nanbeige LLM Lab has launched Nanbeige4.3-33B, a 33B-parameter open-source agentic coding model available on Hugging Face under the Apache 2.0 license. The model is designed for coding and office-style collaboration tasks, with a 524,288-token context window. It outperforms 17 out of 20 benchmarks against other open models near 30B parameters. The model is part of the Nanbeige4.3 series, with the lab’s strategy focusing on building a strong teacher model first before distilling skills into smaller models. It uses a standard LlamaForCausalLM architecture without expert routing, and supports three chat controls for reasoning and tool calling. The model was compared with Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, with self-reported scores showing some gaps. Nanbeige’s Base model outperforms Gemma4-31B-Base on multiple tasks, including MMLU-Pro and GPQA. The model card includes inference recipes and recommended settings, though no Hugging Face provider hosts it at publication. The model requires about 66 GB of BF16 weights and 112 GB of KV cache for a full sequence.