Nanbeige releases Nanbeige4.3-33B, an open model for agentic coding

Nanbeige has released Nanbeige4.3-33B, an open-weight language model built around 33 billion parameters, released under the Apache 2.0 license. The company describes it as the flagship teacher model in a planned Nanbeige4.3 family, meant to anchor a later distillation pipeline that will train smaller models, including a planned 3B version, from its agentic skills. It supports a context window of up to 512K tokens and ships with chat templates for reasoning and tool-calling.

Nanbeige says the model targets agentic coding and professional cowork tasks, and the model card lists comparisons against three other roughly 30B-parameter open models (Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B) across benchmarks covering coding, planning, and office-style agent work. Nanbeige reports it leads on most of these benchmarks, including strong scores on coding tests like Terminal-Bench and SWE-Atlas-QnA, and claims it outperforms the baseline on most of the 35 occupations covered by JobBench.

All of these benchmark results are self-reported by Nanbeige on the model's Hugging Face page, run using the company's own evaluation harness and hyperparameters, so they haven't been independently verified. The model can be run through Transformers, SGLang, or vLLM.

Why it matters

It's another sizable open-weight model aimed squarely at coding agents, giving developers a free alternative to try for agentic workflows.

Source: Nanbeige on Hugging Face