Skip to content
daily-hour-news·

🤖Cohere's North Mini Code: 30B Coding Agent on One H100

TL;DR

Cohere open-sourced North Mini Code, a 30B-parameter MoE coding model that runs on a single H100 under Apache 2.0. It ships a 256K context window and hit 210 tokens per second in Artificial Analysis testing.

Cohere open-sourced North Mini Code, a 30B-parameter MoE coding model that runs on a single H100 under Apache 2.0. It ships a 256K context window and hit 210 tokens per second in Artificial Analysis testing.

Cohere's North Mini Code: 30B Coding Agent on One H100 — daily-hour-news

Key Points

1

30B MoE with 3B active params per token; Apache 2.0 license

2

256K context window, 64K max generation length

3

Runs on a single NVIDIA H100; 210 tok/s output (8th of 127 open models)

4

Built for sub-agent orchestration, code review, and terminal work

Why It Matters

A capable coding agent that fits on one GPU lets teams self-host instead of renting frontier APIs, a real cost and privacy lever for enterprises.

Quick Facts

CohereNorth Mini Codeopen weightscoding agentsMoEApache 2.0local LLM

Frequently Asked Questions

Why does this matter?

A capable coding agent that fits on one GPU lets teams self-host instead of renting frontier APIs, a real cost and privacy lever for enterprises.

What happened?

Cohere open-sourced North Mini Code, a 30B-parameter MoE coding model that runs on a single H100 under Apache 2.0. It ships a 256K context window and hit 210 tokens per second in Artificial Analysis testing.

Comments

Subscribe to join the conversation...

Be the first to comment

Enjoyed this article?

Get it daily. 7am. Free. Reads in 5 minutes.