Cohere North Mini Code

North Mini Code 1.0 is a 30B parameter decoder-only sparse MoE model with 3B active parameters, designed specifically for agentic software engineering, terminal workflows, and code generation. It features 49 layers, 128 experts with 8 activated per token, and a 3:1 interleaved attention pattern combining 4,096-token sliding-window RoPE attention with global attention. The model uses SwiGLU FFNs, sigmoid-based expert routing, and a 2,048-dimensional hidden size. Post-trained with SFT and RLVR, it supports tool use, 256K context, and 64K output lengths, targeting efficient, reliable coding agents.