Introducing Cohere's first open-source coding model: North Mini Code Small & efficient, designed for agentic performance and built for community input.

Nice! Great to see another open source model released. 🙌

For the community 🤝

Small: 30 billion parameters, 3B active. Efficient: Benchmarks to 33.4 on the Artificial Analysis Coding Index, competitive among similar sized models. Open Source: Apache 2.0 license so developers can experiment, test, and build their way. Learn more:

We encourage developers to share their builds with us and give feedback to shape future iterations. Let’s shape the future of sovereign AI together. Download:

Did you literally just post a six-minute launch video where it fails edit tool calls for half of the video?

Open source model over closed source. Any day.

banger after banger after banger thanks canada bros 🙏

How do these 2 heads work in the same office without any conflict

30B total / 3B active is a wild MoE config. Speed of a 3B model, but you still need the VRAM of a 30B to hold all 128 experts. Local VRAM is the real bottleneck here, but the TPS should be ridiculous.

congrats!

Wow! This is another cool release!

codehere

So happy we at @daytonaio could be a small part of this 🙏

tuesday launch day is always such a great day

congrats team on the model launch!

@DJLougen you’re up 😈

I mean, I would very much like Cohere to succeed, but why should I use this instead of qwen3.6 35b again?

nice to see another moe at 30b/3b form factor. apache 2.0 is the real win. would love to see how agentic coding shakes out vs qwen3 on the same hardware, that 33.4 benchmark number only tells part of the story

Congrats on the launch! So happy to have been a part of this process 🚀

I liked the honest bit about usability. Could be nice for some offline robots as backup I think

Great work. It's a great time to be alive in the age of AI.

It’s going to be less resource intensive to run models locally than GTA-6

Really cool and a big win for everyone! Would love it for the data to be open too. Then it can truly be called open source, like @allen_ai’s LLMs.

Awesome, congrats and thanks for sharing!

Nice size to target, be also good to get the ~8b locked down too

Simple and efficient models can be very helpful.

Interesting they're calling it "Mini" — curious how the agentic performance holds up vs Qwen-Coder 2.5 at similar parameter counts.

Small, efficient and open source for agentic coding is a solid combo. How does it actually benchmark against the other small coding models?

Making Canada proud. LFG

Gave it a try. Pretty fast on my m5 max MacBook Pro. It did ok but for some reason it was determined to use Deno instead of Node.js? Also couldn’t handle pnpm… performance is good, but reasoning and output need to be improved.

Awesome work, team. Appreciate you guys contributing to the ecosystem!

Make a bigger real one and deploy via Cursor Canadian coding model that everyone can use and is near-SOTA

Exciting to see open-source coding models becoming more efficient and accessible. North Mini Code highlights that strong agentic performance doesn't always require massive models—optimization and community-driven innovation matter just as much. Great work by the Cohere team. 🚀

Nice