A sparse Mixture-of-Experts model that activates a subset of its 8×7B experts per token to deliver efficient, high-performance instruction following. The MoE design gives it the capacity of a much larger dense model while keeping compute and memory usage moderate. It performs well across structured tasks, coding, and dialog, making it popular in production systems seeking MoE-style cost savings. It serves as a robust, open-weight backbone for many instruction-driven use cases.
