XGen is a family of open-source large language models from Salesforce AI Research, designed around an 8K sequence length for long-context understanding. The repository collects the model releases and code behind the XGen series. It has about 728 stars on GitHub and saw updates in 2026.
Salesforce introduced XGen as its open contribution to the LLM ecosystem, with models trained for tasks where context length matters: long documents, extended conversations and multi-step instructions. The 8K token window was a headline feature at release, and the family includes instruction-tuned variants aimed at practical assistant-style use. Released under the Apache-2.0 license, the weights and supporting code are available for researchers and developers to download, fine-tune and deploy in their own environments. XGen remains a useful reference point in the open-model landscape, particularly for teams evaluating long-context architectures from established research labs.
The best-known release in the family is XGen-7B, published in base and instruction-tuned versions and benchmarked against contemporary open models on summarization, question answering and instruction-following tasks. The repository remains the canonical home of the series, collecting its checkpoints, release notes and usage documentation in one place.
XGen is free and open source under the Apache-2.0 license. There is no hosted subscription from this repository; you download the models and run them on your own infrastructure, paying only for your compute.
XGen is for developers and researchers who want an openly licensed, long-context language model from a major research lab. It fits teams building document-heavy applications, or anyone benchmarking open models for summarization, question answering and instruction following.