--- title: "Cohere's North Mini Code Now Runs Locally on a Mac at 67.6% SWE-Bench" subtitle: "Cohere's 30B MoE coding model now runs locally on a Mac via a 4-bit quantized build, with free API access on OpenRouter and native Ollama support" url: "https://alphasignal.ai/news/cohere-s-north-mini-code-now-runs-locally-on-a-mac-at-67-6-swe-bench" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Cohere" company_url: "https://alphasignal.ai/company/cohere" published: "2026-06-18T18:09:46.000Z" updated: "2026-06-18T19:18:28.251Z" type: "News" topics: - "Development" - "Open Source" subtopics: - "Code Agents" upvotes: 790 read_time: "5 min" source_url: "https://huggingface.co/CohereLabs/North-Mini-Code-1.0-w4a16" comments: 0 --- # Cohere's North Mini Code Now Runs Locally on a Mac at 67.6% SWE-Bench > Cohere's 30B MoE coding model now runs locally on a Mac via a 4-bit quantized build, with free API access on OpenRouter and native Ollama support Cohere just made its first open-source coding model significantly easier to run. [North Mini Code 1.0](https://huggingface.co/CohereLabs/North-Mini-Code-1.0-w4a16) launched a week ago as a 30B-parameter Mixture-of-Experts model, and today Cohere is shipping three updates that push it onto consumer hardware: a 4-bit quantized build that fits on a Mac, free access through the OpenRouter API, and native support on Ollama. ## What North Mini Code actually is North Mini Code is the first model in Cohere's new "North" family, specifically designed and trained for agentic software engineering tasks. It routes each query to a small subset of specialized "expert" networks within the larger model, 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. It is a meaningful departure: Cohere's prior models, Command R and Aya, were not coding-focused, and the company had not previously shipped an open-weight coding model. The Apache 2.0 license is about as permissive as open-source gets, meaning companies can modify, deploy, and even commercialize the model without licensing headaches. ## Three new ways to run it The headline update is the W4A16 quantized checkpoint. This uses 4-bit weights with 16-bit activations, shrinking the memory footprint to roughly 18-20GB , small enough to run on a Mac Studio via MLX. Cohere co-founder Nick Frosst demoed it on a Mac Studio via MLX at roughly 20 GB of RAM. The two other updates are: - **OpenRouter:** North Mini Code is available on OpenRouter for free, letting you call it via a standard OpenAI-compatible API without managing any infrastructure. - **Ollama:** The model is now in the Ollama library, so you can pull and run it locally with a single command alongside tools like Codex and OpenClaw. ## The quantization is smarter than it looks The W4A16 format isn't a naive weight dump. Cohere applies NVFP4 quantization only to the MoE expert layers , which hold the bulk of the parameters , while keeping attention, the dense layer, and the router at higher precision. To recover quality lost from quantization, they use **Quantization-Aware Distillation (QAD)** , a technique where the quantized model is trained to match the outputs of the full-precision model, rather than just being post-hoc compressed. The result is greater than 99% overall accuracy recovery across evaluations. Because only weights are quantized (not activations), it doesn't require next-generation Blackwell GPUs and runs fine on Hopper and Ada hardware. ## Benchmark numbers worth knowing 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). ![Benchmark comparison table showing North Mini Code performance vs Qwen3.6, Poolside XS.2, Gemma4, and Devstral Small 2 across SWE-Bench, Terminal-Bench, SciCode, and LiveCodeBench](https://cdn-uploads.huggingface.co/production/uploads/62668f725fb8d521d94d8451/xR7kZ3X9RKEZrbgD6hpG1.png) On SWE-Bench Verified, the model resolves 67.6% of tasks at pass@1 using the SWE-Agent harness. In Cohere's internal tests, North Mini Code reached up to 2.8x higher output throughput at identical concurrency and hardware, and a 30% edge in inter-token latency compared to Devstral Small 2. The honest caveat: for pure coding quality, frontier models like GPT-4 and Claude still lead on most benchmarks. But North Mini Code offers something they can't: full local deployment with no API costs and complete data privacy. ## How it was built The post-training pipeline is where things get interesting. Cohere used a two-stage supervised fine-tuning (SFT) approach followed by reinforcement learning with verifiable rewards (RLVR). The first SFT stage used a broad mix where code data made up 70% of trainable tokens. The second stage narrowed to 4.5 billion tokens of only agentic and reasoning-driven samples, all verified as executable and correct. The training data spanned over 70,000 verifiable tasks across roughly 5,000 unique repositories, deduplicated against SWE-Bench to avoid leakage. One key design choice: rather than optimizing a single agent harness (like SWE-Agent), Cohere trained across multiple scaffolds , SWE-Agent, mini-SWE-Agent, and OpenCode. This cross-harness training yielded a 10% gain on OpenCode evaluations while maintaining SWE-Bench performance, and the model generalized to harnesses it wasn't explicitly trained on. The RLVR stage then improved pass@1 on Terminal-Bench v2 by 7.9 percentage points and SWE-Bench by 3.0 points over the SFT-only baseline. The RL training itself used an asynchronous setup to handle the highly variable length of coding agent rollouts. A vLLM sidecar continuously generates rollouts while the trainer updates weights every few steps, avoiding the GPU idle time that a synchronous loop would cause. The loss function is CISPO , a token-level importance-sampling objective that gives more gradient signal to long agentic traces, where the actual credit-assignment information lives. ## Where it fits in your stack The model is designed for agentic software engineering (repo-level code changes inside harnesses like SWE-Agent and OpenCode) and terminal-based agents that drive shell tools end-to-end across multi-turn tasks. For an enterprise team where compliance requires that no proprietary data leaves the building, this is the kind of model that actually gets past legal review. Here's how to get it running locally in three different ways: - **Ollama:** `ollama run north-mini-code-1.0` , simplest path, no config needed - **MLX (Mac):** Uses the W4A16 quantized weights, ~20GB RAM required - **vLLM (server):** Requires the vLLM main branch and Cohere's `cohere_melody` library for correct tool-call parsing - **OpenRouter API:** Free tier available at [openrouter.ai/cohere/north-mini-code:free](https://openrouter.ai/cohere/north-mini-code:free) - **Hugging Face Transformers:** Use the BF16 weights (`CohereLabs/North-Mini-Code-1.0`); the W4A16 checkpoint is not compatible with Transformers The context window is 256K tokens with a 64K max output length , generous enough for whole-repo context. The model supports interleaved thinking and works best when it is turned on, so make sure to enable it in whatever harness you're using. The weights are on [Hugging Face](https://huggingface.co/CohereLabs/North-Mini-Code-1.0-w4a16) in BF16, FP8, and the new W4A16 format, all under Apache 2.0. ## Sources - [X post by @cohere](https://x.com/cohere/status/2067671125073576382) - [CohereLabs/North-Mini-Code-1.0-w4a16 ยท Hugging Face](https://huggingface.co/CohereLabs/North-Mini-Code-1.0-w4a16) - [Introducing North Mini Code: Cohere's First Model For Developers](https://huggingface.co/blog/CohereLabs/introducing-north-mini-code) - [Cohere North Mini Code: An Open 30B Agentic Coding Model](https://www.digitalapplied.com/blog/cohere-north-mini-code-open-source-30b-coding-model-release) - [North Mini Code | Cohere Docs](https://docs.cohere.com/docs/north-mini-code-1.0) - [Meet North Mini Code - MarkTechPost](https://www.marktechpost.com/2026/06/11/meet-north-mini-code-coheres-30b-open-weight-mixture-of-experts-model-with-3b-active-parameters-for-agentic-coding/)