Liquid AI 8B-A1B MoE

Liquid AI has released the 8B-A1B MoE, a sparse mixture-of-experts language model trained on 38 trillion tokens. With 8 billion total parameters but only 1 billion active per token, it achieves high efficiency in both training and inference. The model is designed to deliver strong performance while reducing computational costs, making it suitable for deployment in resource-constrained environments. It builds on Liquid AI's previous LFM series and represents a significant step in scaling MoE architectures. The model is available under a permissive license, encouraging community adoption and further fine-tuning. Early benchmarks suggest it competes favorably with other open-weight models of similar active parameter count.

Liquid AI officially announced the model on their blog, detailing its training on 38T tokens and sparse MoE design, generating discussion on Hacker News.