--- title: "Cohere's North Mini Code Goes Open Source and Beats Models 4x Its Size" subtitle: "Cohere's first open-source coding model, North Mini Code, is a 30B MoE with only 3B active parameters — and the community already ported it everywhere in 48 hours." url: "https://alphasignal.ai/news/cohere-s-north-mini-code-goes-open-source-and-beats-models-4x-its-size" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Cohere" company_url: "https://alphasignal.ai/company/cohere" published: "2026-06-11T21:29:02.000Z" updated: "2026-06-16T19:22:11.297Z" type: "News" topics: - "Open Source" - "Development" subtopics: - "Code Generation" upvotes: 695 read_time: "7 min" source_url: "https://north-mini-code-docs.docsalot.dev/getting-started/overview" comments: 0 --- # Cohere's North Mini Code Goes Open Source and Beats Models 4x Its Size > Cohere's first open-source coding model, North Mini Code, is a 30B MoE with only 3B active parameters — and the community already ported it everywhere in 48 hours. [North Mini Code](https://cohere.com/blog/north-mini-code) is Cohere's first open-source model built specifically for developers, and the community response has been unusually fast. Within 48 hours of release, independent contributors had already shipped llama.cpp support, MLX support for Apple Silicon, community GGUF quants, and a CLI task manager built on top of the model , all before Cohere's own official tooling had fully landed. ## A coding model that punches above its weight class North Mini Code 1.0 is a 30 billion parameter Mixture-of-Experts model, launched under the Apache 2.0 license and freely available on Hugging Face. The key number is not 30B, though. The model routes each query to a small subset of specialized expert networks, with 30 billion parameters total but only 3 billion active at any given time, keeping inference costs dramatically lower than a dense 30B model would require. North Mini Code is a sparse mixture-of-experts model with 128 experts, of which 8 activate per token. This is the same architectural trick that made Mixtral and Qwen3 so efficient , you get the capacity of a large model at the compute cost of a small one. The model supports a context length of 256K tokens and can generate outputs up to 64K tokens. ## Built for agents, not just autocomplete North Mini Code is the first model in Cohere's new family of models, and is specifically designed and trained for agentic software engineering tasks. The distinction matters. Most coding models are fine-tuned general-purpose models. This one was built from the ground up for multi-step agentic workflows where the model needs to call tools, read terminal output, and iterate. It is built for agentic workflows, including understanding and orchestrating sub-agents, mapping systems architecture, and running code reviews. It has integrated tool-use capabilities and supports interleaved thinking, which Cohere says improves performance across multi-step agentic work. The training pipeline reflects this focus. Cohere trained the model through two stages of supervised fine-tuning followed by reinforcement learning with verifiable rewards across more than 70,000 verifiable tasks spanning approximately 5,000 repositories, deduplicated against SWE-Bench. The RL stage (RLVR) rewarded the model for actually solving tasks, not just generating plausible-looking code. ## Harness robustness: the hidden differentiator One of the more interesting technical choices Cohere made is training across multiple agent scaffolds rather than optimizing for one. Real-world code agents use wildly different interfaces , some expose a rich CLI with specialized commands, others give the model a single bash tool and raw shell output. Rather than optimizing against a single agent scaffold, Cohere trained across three: SWE-Agent uses a rich CLI with specialized commands; Mini-SWE-Agent uses a single bash tool with raw shell output; and OpenCode uses individually typed tools returning structured JSON. Cohere reports a 10 percentage point gain on OpenCode evaluation from the multi-harness approach while maintaining SWE-Agent performance. In practice, this means the model should work well whether you drop it into OpenCode, a custom bash-based agent, or a structured JSON tool-calling setup. ## The numbers On Artificial Analysis' Coding Index, North Mini Code achieves a score of 33.4, outperforming Qwen3.5 (35B-A3B), Gemma 4 (26B-A4B), Devstral Small 2 (24B Dense), and even substantially larger models such as Nemotron 3 Super (120B-A12B), Mistral Small 4 (119B-A6B), and Devstral 2 (123B). Speed is also a selling point. Cohere reports North Mini Code at 2.8x higher output throughput and approximately 30% lower inter-token latency than Devstral Small 2, measured under identical hardware configurations. Those are vendor internal tests, not independently confirmed. One caveat worth noting: Artificial Analysis measured approximately 3x the output tokens of comparable models, which benchmark rankings skip but high-volume pipelines cannot ignore. ## The community shipped faster than the docs This is where the story gets interesting. Cohere's tweet thread two days after launch reads less like a company announcement and more like a changelog written by the community. Here is what landed on its own: - **llama.cpp support:** The model uses the `cohere2moe` architecture, which is not in a stock llama.cpp release yet. Until llama.cpp PR #24260 is merged, users can build from that PR branch , and once it lands, the same GGUFs will run on stock llama.cpp with no re-download. - **MLX support:** Day-zero support for Apple Silicon via MLX was contributed by community member Prince Canuma, meaning Mac users could run the model locally almost immediately after release. - **Unsloth GGUFs:** Run North Mini Code locally with a llama.cpp setup, Ollama, or vLLM, with quants from 9GB to full BF16. The quantization range means you can run a usable version on a consumer GPU with ~10GB of VRAM. - **oMLX integration:** Community member michaelasper on Hugging Face built support directly into the oMLX project. - **Taskflow CLI:** Andrew Girgis built a full CLI task manager powered by North Mini Code, allowing users to add, manage, and organize tasks with titles, descriptions, priorities, and tags. - **Community docs:** A community member built an entire [third-party documentation site](https://north-mini-code-docs.docsalot.dev/getting-started/overview) covering deployment paths, quickstart guides, and agentic coding workflows. ## Who wins here The clearest winner is any team that wants a capable coding agent they can self-host. You can download the weights and run on a single H100 at FP8, or approximately 20GB of RAM on Apple Silicon via MLX. That is a meaningful bar , a single H100 is accessible to most teams, and Apple Silicon support means individual developers can run it on a laptop. The release is positioned around "sovereign" AI , the idea of running capable models on your own terms. Small, efficient coding models let teams self-host without large GPU clusters, and North Mini Code targets that gap directly. For enterprises with data residency requirements or teams that simply do not want to send proprietary code to an external API, this is the pitch. The model is strong and narrow , an Agentic Index of 21.7 shows it underperforms on non-coding agentic tasks. The production pattern is a frontier orchestrator with North Mini Code handling code execution on cheap fixed hardware. Think of it as a specialized sub-agent, not a general-purpose assistant. ## What to watch The release comes less than three weeks after Cohere launched Command A+, its previous model. Command A+ received a score of 37 on the Artificial Analysis Intelligence Index, and that cadence suggests the company is accelerating its release schedule. North Mini Code is explicitly labeled as the first in a new "North" family, so more models in this line are likely coming. The immediate practical blocker is llama.cpp. Until PR #24260 is merged into the main branch, users need to build llama.cpp from that PR branch to load these files. That is a one-time compile step, but it is friction that will disappear once the PR lands. If you want to try the model today, the fastest path is Ollama with the Unsloth GGUFs, or the [hosted Hugging Face Space](https://huggingface.co/spaces/CohereLabs/North-Mini-Code-1.0) Cohere provides for zero-setup evaluation. For production agentic use, vLLM is supported, though it requires installing Cohere's `cohere_melody` library alongside the vLLM main branch for accurate response parsing. The open-source coding model space is now genuinely crowded , Mistral Devstral, Qwen3's coding variants, and others are all competing for the same workload. What North Mini Code adds to that mix is a model trained explicitly for harness diversity and terminal-based agentic tasks, with a community that clearly moved fast to make it accessible everywhere that matters. ## Sources - [X post by @cohere](https://x.com/cohere/status/2065184558572532143) - [Introducing North Mini Code: Cohere's First Model For Developers](https://huggingface.co/blog/CohereLabs/introducing-north-mini-code) - [North Mini Code: Agentic Coding Model for Developers | Cohere](https://cohere.com/blog/north-mini-code) - [unsloth/North-Mini-Code-1.0-GGUF · Hugging Face](https://huggingface.co/unsloth/North-Mini-Code-1.0-GGUF) - [Cohere open-sources a coding agent that runs on a single H100 | VentureBeat](https://venturebeat.com/technology/cohere-open-sources-a-coding-agent-that-runs-on-a-single-h100) - [North Mini Code: Cohere's small coding-focused MoE model | Artificial Analysis](https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model)