Cohere AI released North Mini Code, its first open‑weight coding model aimed at software engineers. The model is a 30‑billion‑parameter mixture‑of‑experts (MoE) transformer that activates only 3 billion parameters per token, keeping compute low while preserving capacity. It is optimized for three core tasks: code generation, agentic software engineering, and terminal operations. The context window stretches to 256 K tokens with a maximum output of 64 K tokens. Minimum hardware to run the model is a single H100 GPU in FP8 precision; the weights are available under an Apache 2.0 license on Hugging Face, Cohere API, Model Vault, and OpenRouter.
Performance benchmarks show a score of 33.4 on the Artificial Analysis Coding Index, placing it competitively among similarly sized models. In internal tests, North Mini Code delivers up to 2.8 × higher output throughput and a 30 % improvement in inter‑token latency compared with Devstral Small 2, while time‑to‑first‑token remains comparable.
Typical use cases include sub‑agent orchestration (e.g., one agent writes tests while another fixes code), systems architecture mapping (reading a repo to sketch service interactions before a refactor), and code reviews (scanning diffs for issues such as unguarded null dereferences). Terminal tasks like listing files, running builds, and parsing error output also fit naturally.
To get started quickly, install the latest Transformers library from source, then load the model and tokenizer:
pip install “git+https://github.com/huggingface/transformers.git”
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, device_map=”auto”)
prompt = “Write a Python function that checks if a string is a palindrome.”
inputs = tokenizer.apply_chat_template([{“role”:”user”,”content”:prompt}],
add_generation_prompt=True,
return_tensors=”pt”).to(model.device)
output = model.generate(**inputs,
max_new_tokens=1024,
do_sample=True,
temperature=1.0,
top_p=0.95)
print(tokenizer.decode(output[0][inputs.shape[-1]:]))
For serving, vLLM works with the cohere_melody library:
uv pip install “git+https://github.com/vllm-project/vllm.git”
uv pip install “cohere_melody>=0.9.0”
vllm serve CohereLabs/North-Mini-Code-1.0 \
-tp 2 \
–max-model-len 320000 \
–tool-call-parser cohere_command4 \
–reasoning-parser cohere_command4 \
–enable-auto-tool-choice
Quantized versions are also available for Ollama, LM Studio, and llama.cpp, and you can try the model via Cohere’s free OpenCode playground or a hosted Hugging Face Space.