Nanbeige publishes a 33B agent model under Apache 2.0

The model card describes a 512K context window and configurable tool use, with smaller distilled agents still planned.

Nanbeige has published Nanbeige4.3-33B, an open-weight model positioned for coding and professional agent tasks. Its Hugging Face repository carries an Apache 2.0 licence and identifies the model as a 33-billion-parameter teacher for a planned family of smaller agents.

Repository metadata shows it was created on October 24 and updated on October 25. Those timestamps establish public repository activity, rather than a separate announcement date.

A configurable agent model

The model card lists a maximum context of 524,288 tokens and options for reasoning and tool-call formatting. It provides deployment examples for Transformers, SGLang and vLLM. The team presents comparisons with other models around the same parameter class; those are developer-reported evaluations.

A long context limit does not by itself show how accurately a model can locate a detail within a large repository or retain instructions across many actions. Those behaviours need checking in the application that will use it.

Tool use has explicit settings

The model’s chat template can switch reasoning on or off and decide whether earlier reasoning is retained across turns. Its authors recommend retaining that context for multistep tool and office tasks, while disabling it for general chat. Tool calls support XML and JSON formatting, with the card recommending XML.

Nanbeige describes the 33B model as a teacher for subsequent smaller models, including a planned 3B version. That future work is distinct from the weights available now. The release gives operators control over where inference runs, but the surrounding agent still needs to handle malformed calls and check whether a tool actually completed the intended task.

Sources

- [Nanbeige model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) — Nanbeige