HexaBee-4B is a translation model introduced in a new arXiv preprint. It starts from Gemma 3 4B IT and is adapted in three stages, beginning with continued pretraining on Wikipedia. The abstract also mentions a joint stage, though the full description is cut off in the source.

According to the abstract, the model is designed for six-language translation and multilingual answering, with bidirectional supervision to help preserve meaning. The authors frame the task as requiring a model to keep meaning intact while producing text in the requested language.

The paper is available on arXiv under identifier 2610.16361. As the source is a single abstract, no independent comparisons or external evaluations are reported here.