Meet ‘North Mini Code’: Cohere’s 30B Open-Weight Mixture-of-Experts Model With 3B Active Parameters for Agentic Coding
This week, Cohere AI team shipped its first developer-facing coding model named ‘North Mini Code‘. ‘North Mini Code’ is open-weight and focused at software engineers. It is a mixture-of-experts (MoE) model with 30B total parameters. Only 3B of those parameters activate per token.
The release is positioned around “sovereign” AI. The idea is simple: run capable models on your own terms. Small, efficient coding models let teams self-host without large GPU clusters. North Mini Code targets that gap directly.
North Mini Code
North Mini Code is a 30B-A3B parameter model. The A3B stands for three billion active parameters per forward pass. Cohere optimized it for three jobs: code generation, agentic software engineering, and terminal tasks. The model is text-in, text-out. There is no image or video input.
The context window is 256K tokens. Maximum output length is 64K tokens. Cohere lists a minimum hardware bar of one H100 at FP8. Weights ship under Apache 2.0 on Hugging Face. You can also reach it through the Cohere API, Model Vault, and OpenRouter.
The Architecture
North Mini Code is a decoder-only Transformer with sparse MoE layers. Its attention interleaves two types in a 3:1 ratio. Sliding-window attention uses RoPE for positions. Global attention uses no positional embeddings at all. The feed-forward block holds 128 experts. Eight experts activate per token. Each expert is an FFN with SwiGLU activation.
The router applies a sigmoid before top-k selection. A single dense layer sits before the sparse layers. That mix keeps active compute small while widening total capacity. Cohere released the weights in BF16.
Post-training ran in two phases. First came two-stage cascaded supervised fine-tuning (SFT). Then came reinforcement learning with verifiable rewards (RLVR). The post-training focused on agentic coding. The model also supports interleaved thinking and native tool use.
Benchmarks
Cohere reports a 33.4 on the Artificial Analysis Coding Index. It describes this as a competitive position among similarly sized models. The company evaluated on SWE-Bench Verified, SWE-Bench Pro, and Terminal-Bench v2. It also used Terminal-Bench Hard, SciCode, and LiveCodeBench v6.
The methodology is specific. SWE-Bench used the SWE-agent harness v1.1.0. Terminal-Bench v2 used a simple ReAct harness with one terminal tool. Terminal-Bench Hard used the Terminus-2 harness. Each benchmark ran with three seeds, then averaged. Sampling used temperature 1.0 and top_p 0.95.
The Speed
In Cohere’s internal tests, North Mini Code reached up to 2.8x higher output throughput. That held at identical concurrency and hardware. It also showed a 30% edge in inter-token latency. Time-to-first-token was closer between the two. Devstral Small 2 kept a slight TTFT lead.
Use Cases With Examples
Cohere built North Mini Code for agentic workflows.
Three patterns stand out in its own framing:
Sub-agent orchestration: A main agent delegates subtasks to helpers. Example: one agent writes unit tests while another fixes failing code.
Systems architecture mapping: The model reads a repository and sketches its structure. Example: tracing how services call each other before a large refactor.
Code reviews: The model scans a diff for problems. Example: flagging an unguarded null dereference before a merge.
Terminal tasks fit the model as well. Example: listing files, running a build, then parsing the output for errors.
Getting Started
The fastest path is Hugging Face Transformers. Install Transformers from source for this model. Recommended sampling is temperature 1.0 and top_p 0.95.
# 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 program to check if a string is a palindrome or not.”
messages = [{“role”: “user”, “content”: prompt}]
# return_dict=True yields a dict (input_ids + attention_mask) so **inputs unpacks cleanly
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors=”pt”,
).to(model.device)
gen_tokens = model.generate(
**inputs,
max_new_tokens=1024,
do_sample=True,
temperature=1.0,
top_p=0.95,
)
# Decode only the newly generated tokens, not the prompt
output = tokenizer.decode(gen_tokens[0][inputs[“input_ids”].shape[-1]:])
print(output)
For serving, vLLM works. You need vLLM main plus Cohere’s melody library. Accurate response parsing depends on it.
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 builds exist for Ollama, LM Studio, and llama.cpp. You can also try the model before downloading. Cohere offers free access through OpenCode and a hosted Hugging Face Space.
Key Takeaways
Cohere’s first coding model, North Mini Code, is a 30B mixture-of-experts that activates just 3B parameters per token.
It runs on a single H100 at FP8, with 256K context and 64K max output.
Weights ship under Apache 2.0, though the Hugging Face card adds a non-commercial note.
Cohere official release reports 33.4 on the Artificial Analysis Coding Index, and up to 2.8x throughput over Devstral Small 2.
Built for agentic coding—sub-agent orchestration, architecture mapping, code reviews with native tool use
Marktechpost’s Interactive Explainer
Cohere · Open-Weight Coding Model
North Mini Code
Cohere’s first developer coding model: a 30B mixture-of-experts that activates just 3B parameters per token, built for agentic software engineering and terminal tasks.
30B total params
3B active / token
256K context
64K max output
1× H100 @ FP8
Snapshot
How it works
Performance
Run it
The model at a glance
Open weights, released June 9, 2026. Text in, text out.
Size
30B total / 3B active
Architecture
Sparse MoE (decoder-only)
Min hardware
1× H100 @ FP8
License
Apache 2.0 see note
Context window · drag to explore
128K tokens
a mid-size codebase
8K64K output cap256K max
Relatable sizes are approximate. The exact limits are 256K context and 64K maximum generation.
Optimized for
Code generation
Agentic software engineering
Terminal tasks
Agentic use cases
Sub-agent orchestration
Systems architecture mapping
Code reviews
License note: Cohere’s blog states Apache 2.0. The Hugging Face card adds an acceptable-use addendum and a non-commercial note. Check both before deploying.
The forward pass
Tap any stage to see what it does. The MoE block is where sparsity happens.
Input tokens
→
Dense layer
→
Attention (3:1)
→
MoE block
→
Output
Input tokens
Text is tokenized and fed to a decoder-only Transformer. The model is text in, text out.
Try the router
Each MoE block holds 128 experts. The router selects 8 per token. Route tokens and watch coverage grow.
Coral = the 8 experts firing now. Peach = experts used earlier in the run. Hover a square to inspect.
8 / 128 experts
6.25% of experts run per token, so compute stays small.
Unique experts used0 / 128
Tokens routed0
Route a token →
Auto
Reset coverage
Reported performance
Figures are from Cohere. Independent runs on your own workload still matter.
0
Artificial Analysis Coding Index
0
Output throughput vs Devstral Small 2
0
Better inter-token latency
Throughput
Inter-token latency
Higher is better
Time-to-first-token was closely matched, with Devstral Small 2 holding a slight edge.
Benchmarks: SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench v2, Terminal-Bench Hard, SciCode, LiveCodeBench v6. Harnesses: SWE-agent v1.1.0 (SWE-Bench), a ReAct harness with one terminal tool (Terminal-Bench v2), Terminus-2 (Terminal-Bench Hard). Each run used 3 seeds, averaged, at temperature 1.0 and top_p 0.95.
Quickstart
Hugging Face Transformers, installed from source. Recommended sampling: temperature 1.0, top_p 0.95.
# Install Transformers from source, then:
from transformers import AutoTokenizer, AutoModelForCausalLM
mid = “CohereLabs/North-Mini-Code-1.0”
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, device_map=“auto”)
msgs = [{“role”: “user”, “content”: “Write a Python palindrome checker.”}]
inputs = tok.apply_chat_template(
msgs, add_generation_prompt=True,
return_dict=True, return_tensors=“pt”,
).to(model.device)
out = model.generate(**inputs, max_new_tokens=1024,
do_sample=True, temperature=1.0, top_p=0.95)
print(tok.decode(out[0][inputs[“input_ids”].shape[-1]:]))
Serve with vLLM (+ cohere_melody)
Trained for OpenCode
Native tool use + interleaved thinking
Quantized: Ollama, LM Studio, llama.cpp
Also on Cohere API, Model Vault, OpenRouter
Check out the Model weights and Technical details. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
