North Mini Code

North Mini Code

An open-weights research model from Cohere for writing code, agentic software engineering and terminal tasks. A 30B model with 3B active, 256K context. For developers who run a coding model themselves.

Made by CohereLabs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: North Mini Code is a coding model from Cohere and Cohere Labs, shared with open weights for research. It writes code, works as a software engineering agent and handles terminal tasks. It has 30B parameters in total with 3B active at a time. The context length is 256K, with a 64K maximum output.

Who it suits: It suits developers and researchers who want a coding model they can run on their own hardware, including for agent-style work in a terminal. It takes text in and gives text out.

Title: Who uses it
Industry | Who uses it | How they use it | Example
Software | Application developers | Generate and review code with a local model | Ask for a Python program that checks a string
AI tooling | Agent builders | Power a coding or terminal agent | Run software engineering tasks through an agent
Research | Evaluators | Test coding benchmarks with the open weights | Reproduce agentic coding results

People who need image input can skip it, since the card lists text only.

Where to use it:

• Python with the transformers library
• OpenCode, which the card names for trying the model
• Cohere's hosted Hugging Face Space, to try it before downloading
• Windows, macOS or Linux, to download the files

Title: Top jobs
Use case | Who | What they get back
Code generation | Developers | Programs written from a plain description
Agentic software engineering | Agent builders | Multi-step fixes in a repository setting
Terminal tasks | Developers | Commands and steps for command-line work

How to set it up:

• Python on your computer
• The huggingface_hub package
• A transformers install from the source repository that has this model's changes
• About 61 GB of disk for BF16, or about 15.2 GB at 4-bit by estimate

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Install transformers from its source code, which carries the changes this model needs.
4. Load the model and generate with the recommended sampling settings.

pip install -U huggingface_hub
hf download CohereLabs/North-Mini-Code-1.0

How to use it: Load the tokenizer and model, format your request with the chat template, and generate. The card recommends a temperature of 1.0 and top_p of 0.95.

# pip install transformers
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "CohereLabs/North-Mini-Code-1.0"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Write a python program to check if a string is a palindrome or not."

# Format message with the North-Mini-Code-1.0 chat template
messages = [{"role": "user", "content": prompt}]
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
)

gen_tokens = model.generate(
    **input_ids,
    max_new_tokens=1024,
    do_sample=True,
    temperature=1.0,
    top_p=0.95
)

gen_text = tokenizer.decode(gen_tokens[0])
print(gen_text)

The card also shows the transformers pipeline abstraction as another way to run it.

Title: Benchmarks the card names
Benchmark | Kind
SWE-Bench Verified and SWE-Bench Pro | Agentic coding
Terminal-Bench v2 and Terminal-Bench Hard | Agentic terminal tasks
SciCode and LiveCodeBench v6 | Complex code generation outside tool use

Example: You send the prompt asking for a Python program that checks whether a string is a palindrome. The model writes the code as text, and the script prints it. The card shows no printed output, so run the code before you trust it.

For newcomers and experts: If you are new, try the hosted Space first, then download the weights if you like the results. A first small job is a short function in a language you know, so you can judge the code yourself.

For experts, the card lists 3 seeds for each benchmark and recommends sampling at temperature 1.0 with top_p 0.95. The agentic coding runs used SWE-Agent and a simple ReAct setup with a terminal tool. The scores sit in a figure that this text does not reproduce.

Install transformers with this model's support
Load the tokenizer and model
Format the request with the chat template
Generate code with temperature 1.0

Good to know:

• Research release: the card calls it an open-weights research release.
• Input and output: text only.
• Install note: transformers must come from the source repository with this model's changes.
• License: Apache 2.0.
• Cost to run: none named, and a 30B model needs a lot of memory.
• Data it reaches: it downloads files from Hugging Face, and the source names nothing else.
• Not for: image input, since the model takes text only.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 30.5B
• Weights: about 61 GB in BF16; about 15.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 8,171
• Likes: 587
• License: apache-2.0
• Quantized versions: 39
• Fine-tunes: 5
• Tags: transformers, safetensors, cohere2_moe, text-generation, conversational, chat, code, agent

Links

• Model: https://huggingface.co/CohereLabs/North-Mini-Code-1.0
• Files: https://huggingface.co/CohereLabs/North-Mini-Code-1.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download CohereLabs/North-Mini-Code-1.0

Then follow the usage in its model card. Nothing in this kit runs by itself.