Cohere Open-Sources North Mini Code, a 30B Agentic Coding Model Optimized for Single-GPU Inference

Cohere released North Mini Code on June 9, 2026, positioning it as a 30 billion parameter agentic coding agent. [3] The model is built from scratch for agentic software engineering, including architecture mapping, multi-file code review, terminal-based tasks, and sub-agent orchestration. [3] Under the hood, the architecture is a sparse mixture-of-experts Transformer with 128 experts, activating eight per token. [3]

This design gives the model the effective compute profile of a 3B dense model at inference time despite 30B total parameters. [3] Cohere co-founder Nick Frosst demoed the model running on a Mac Studio via MLX at around 20 gigabytes of RAM. [3] Released under an Apache 2.0 license, North Mini Code is available on Hugging Face in BF16 and FP8 quantized weights, and via the Cohere API. [3] The model features a 256,000 token context window and a 64,000 token maximum generation length. [3] North Mini Code natively supports Transformers versions 5.9.0 and higher, with vLLM and SGLang paths in active development. [3]

Independent testing by Artificial Analysis ranks the model 8th of 127 comparable open-weight models on output speed. [3] The model achieves 210 tokens per second, with a time to first token of 0.25 seconds against a class median of 1.95 seconds. [3] North Mini Code achieved a score of 33.4 on the Artificial Analysis Coding Index. [9]

However, the model generated 75 million output tokens in Artificial Analysis benchmarking against a class median of 25 million. [3] That volume is three times the output of comparable models. [3] The model requires a minimum of one H100 GPU running at FP8 or FP4. [9]

For developers deciding whether to route agentic coding workloads through a managed API or a self-hosted open-source model, North Mini Code emerges as a clear contender in 2026. By offering an Apache 2.0 licensed model that operates with three billion active parameters, Cohere provides a tool that delivers software development performance without demanding extensive hardware. The combination of a 256,000-token context window and a 0.25-second time to first token makes it suitable for complex, rapid-response coding environments. However, because the model produces three times the output tokens of comparable models, teams will need to account for higher throughput during orchestration. What the sources don't address: How the model's extreme verbosity impacts overall memory and compute costs during sustained, multi-turn agentic workflows in production environments.