Intel / article

North Mini Code Cohere S Small Coding Focused Moe Model

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Jun 9, 2026

Context

Cohere released North Mini Code less than a month after Command A+, and it is a much smaller model aimed at coding (artificialanalysis.ai). It is a text-only model with 30 billion parameters in total, of which about 3 billion are active for each , so most of its weights sit idle on any single step. It is released under the Apache 2.0 license.

On Artificial Analysis's Intelligence Index it scores 27.6, above gpt-oss-20B (high) at 24.5 and just below Mistral Small 4 at 27.8, a larger model with 119 billion parameters and 6.5 billion active. Its strength shows on the Coding Index, a weighted average of Terminal-Bench Hard and SciCode, where it scores 33.4: well above GLM-4.7-Flash at 25.9 and a little behind Qwen3.6 35B A3B at 35.2.

Outside coding it falls off. It scores 14% on GDPval-AA, a test of real-world work tasks, and 37% on τ²-Bench Telecom, a customer service tool-use test, for an Agentic Index score of 21.7. It also uses more output tokens to complete the index than most comparable models. Against that, it ran at about 199 output tokens per second on Cohere's API in pre-release testing, faster than several open weights models of similar size and intelligence.

Key takeaways · AI-distilled

North Mini Code scores 33.4 on the Artificial Analysis Coding Index, well above GLM-4.7-Flash at 25.9 and slightly below Qwen3.6 35B A3B at 35.2 among open models in its size class.
Outside coding it is weak: 14% on GDPval-AA and 37% on tau-squared Bench Telecom, giving an Agentic Index score of 21.7.
It is a text-only 30B total, 3B active model under Apache 2.0, served at about 199 output tokens per second on Cohere's API, but it uses more output tokens on the index suite than most comparable models.

Terms in this piece · Glossary

open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source artificialanalysis.ai