XGen provides three model variants: XGen-7B-4K-Base, XGen-7B-8K-Base, and XGen-7B-8K-Inst (instruction-finetuned). The models are trained to handle long sequences up to 8K tokens and can be loaded via the HuggingFace transformers library for auto-regressive text generation. They use the OpenAI Tiktoken tokenizer for subword tokenization.