InclusionAI

A hybrid-reasoning Mixture-of-Experts model built for production-scale AI agents. Features Hybrid Linear Attention architecture with 5 KDA layers for efficient long-range memory and 1 MLA layer for precise attention per stack. With 124B total parameters and only 5.1B active per token, it matches or beats InclusionAI's 1-trillion-parameter flagship on most benchmarks. Native 256K context, designed to scale toward 1M.