North Mini Code 1.0 – 30B Agentic Code Generator with Tool Use

North Mini Code 1.0 is a 30 B‑parameter (3 B active) decoder‑only Transformer released by Cohere Labs. The model is built as a sparse Mixture‑of‑Experts (128 experts, 8 active per token) and features a hybrid attention scheme (sliding‑window + global) with RoPE. It is tuned for code generation, agentic software‑engineering tasks, and terminal‑style interactions, supporting up to 256 K tokens of context and 64 K tokens of generated output.

The model is packaged for the `transformers` library under the **text‑generation** pipeline tag and is distributed as safetensors. Its README highlights a two‑stage fine‑tuning pipeline (SFT followed by RLVR) focused on tool‑use capabilities. Users can invoke functions such as Bash commands via the built‑in chat template and JSON‑schema tool specifications, enabling interleaved reasoning and automatic tool selection. The model can also be served with vLLM (with Cohere’s melody library) or run locally through the OpenCode framework.

Evaluation is presented on several coding benchmarks (SWE‑Bench, Terminal‑Bench, SciCode, LiveCodeBench) using the Swe‑Agent and ReAct harnesses, demonstrating its strength in agentic coding scenarios. The model is released under an Apache‑2.0 license and can be tried directly in a Hugging Face Space before downloading the weights.