We are releasing XLM-V, a multilingual model with a 1 million token vocabulary [Link]. The model is also open-sourced in HuggingFace Transformers.