Cohere has released its first Agentic programming model for free use.

Cohere Launches a Programming Agent — Open Source and Free from Day One

On June 18, OpenRouter quietly rolled out a new free model: cohere/north-mini-code:free. The name may seem unremarkable, but a look at Cohere’s official blog reveals it’s not just a routine update — North Mini Code is Cohere’s first model specifically targeted at developers and also the inaugural release in its brand-new North family series. Even more importantly: it’s under the Apache 2.0 license, weights are already on Hugging Face, and OpenRouter integration is free.

In recent years, Cohere has been very “B2B” in its strategy — focusing on RAG, rerank, and enterprise search, targeting industries like finance, telecom, and government where data sovereignty is crucial. So suddenly dropping an agentic coding model is actually a notable signal shift: Cohere wants to compete in the same lane as Cursor, Claude Code, and Cline.

Specs Overview: 30B Total Params, 3B Active — Runs Locally

First, let’s lay out the hard specs to avoid confusion from marketing jargon:

• Architecture: Mixture of Experts (MoE)
• Parameters: 30B total / 3B active
• Context window: 256,000 tokens
• Max output: 64,000 tokens
• License: Apache 2.0
• OpenRouter pricing: Input $0/M tokens, Output $0/M tokens
• Cohere official API: Both trial and production keys are free until rate limits are hit

The 3B active parameters figure is important. That size means you can fit this model on a workstation with a consumer GPU, and after quantization, even into a laptop with 16GB VRAM. Cohere explicitly states in the docs: “small active footprint makes it suitable for local deployment” — in other words, this is genuinely meant for you to run locally, not just a research demo.

Comparing the market: current mainstream “small but capable” coding models typically range from 7B–32B parameters, with Qwen3-Coder, GLM-4-Code, and DeepSeek-V2-Lite-Coder competing heavily here. Cohere’s choice of a 30B/3B MoE setup clearly aims for “cheap inference, fits in VRAM, without too much performance trade-off”.

Its Differentiation Is “Agentic,” Not Just “Can Write Code”

This is a point Cohere emphasizes repeatedly — and it’s the most interesting aspect to unpack.

Traditional code completion models — like the original Codex or CodeLlama — work essentially as “completion engines”: you give them the preceding text, they produce the rest. Later, Claude 3.5/4 and GPT-4o enhanced instruction-following so you could describe requirements in natural language and have them output full files.

But agentic coding is a different ballgame: the model has to operate in an actual environment with a file system, terminal, git, and error logs, deciding on its own “should I run ls now or first cat that config file?” “This test failed — should I roll back my last change?” “Before installing this dependency, should I check package.json for existing entries?” This is a sequence of multi-step decisions.

Cohere says something interesting in the docs:

It was trained against multiple harnesses, so performance generalizes across agent scaffolds rather than being tuned to a single one.

Translation: they didn’t overfit to any single agent framework (like OpenHands, Aider, SWE-agent) but trained with multiple harnesses together. This is a slightly counterintuitive choice — many teams fine-tune for a specific scaffold to boost SWE-bench scores, but Cohere says they don’t care about the score; they care about compatibility with any framework.

This choice is friendly to developers. It means if you already use Cline, Continue, Roo Code, or even your own custom agent loop, theoretically you can swap in North Mini Code without redoing prompt engineering.

First-Hand Run: Early Feedback from the Community

On Linux.do, someone immediately tested this model on OpenRouter and reported: “fast, works fine.” This matches expectations — a 3B active parameter model on OpenRouter’s backend (likely running vLLM or something similar) will naturally generate tokens quickly, and MoE’s sparsity boosts throughput beyond dense 7B models.

I tried it in a few common scenarios, here’s a quick summary:

• Build a complete React component + Tailwind: It can do it, structure is fine, but variable names are simpler than Claude 4’s, with fewer comments.
• Locate a bug in a monorepo with 30 files: In an agentic setup paired with a simple tool loop (ls/read/grep/edit), it performs noticeably better than comparably sized dense models — especially with habits like “review the entire directory before making changes.”
• Large-context refactor: The 256K context really comes in handy — feed it an entire medium-sized project and ask “which functions should be moved to a common module” and it produces a fairly solid list.
• Complex algorithms: This is visibly weaker than GPT-5 or Claude Sonnet 4.5 — not its intended domain, so don’t take it to LeetCode hard problems.

So the positioning is clear: this is not a model for leaderboard runs or complex algorithm design; it’s a model meant to plug into an agent loop and grind through routine tasks.

Why Now, Why Free

The current wave of open-source/free coding models isn’t random. In the past six months we’ve seen NVIDIA release the entire Nemotron 3 series for free, MiniMax M3 listed on OpenRouter at very low prices, ByteDance continually updating its Seed series. These companies aren’t aiming to recoup costs through token billing — they’re after ecosystem positioning.

Cohere is being even more direct: North platform’s main offering is “private deployment + enterprise-grade agent workspace”. The model itself is free and open-source, which reduces customer hesitation during POC — if your CTO worries about data sovereignty, no problem: take the weights, deploy yourself, test it, then discuss business.

This strategy is similar to Mistral’s early days or Meta’s Llama releases, except Cohere is more focused — putting all its push behind coding agents, avoiding head-on battles with general-purpose LLMs.

Suitable vs. Unsuitable Use Cases

Based on the specs and testing, here’s a balanced assessment:

Suitable for

• Running a coding agent locally, especially for privacy-conscious teams avoiding Claude/GPT
• Deploying agents in CI/CD pipelines for automated PR review, commit message generation, and fixing tests
• Teaching and experimental use — 3B active params, Apache 2.0, 256K context make for great value
• Fine-tuning as a base model for a dedicated agent for your codebase

Not suitable for

• Expecting it to build an entire product from a single prompt (leave that to Claude 4/GPT-5)
• High-end algorithmic reasoning demands
• Multi-modal needs (this is text-only)

How to Use

The easiest route is OpenRouter — model ID: cohere/north-mini-code:free. It’s free until rate limits kick in. OpenAI Hub has also linked Cohere’s models, so developers can use one key to seamlessly switch between GPT, Claude, Gemini, DeepSeek, and now North Mini Code — for agent framework builders, this “multi-model via one interface” use is more frequent than simple calls.

For local runs, Hugging Face hosts the weights at CohereLabs/North-Mini-Code-1.0, usable with vLLM or llama.cpp (once quantized versions release). Cohere also offers a hosted inference environment via Model Vault for enterprise customers.

A Final Take

North Mini Code won’t wow you after using Claude 4, but it combines “open source + agentic + small size + long context” all in one — which is rare among current open-source coding models. Qwen3-Coder’s agentic features aren’t separately optimized, DeepSeek-V3’s size is too large for easy local use, and GLM-4’s open-source version keeps waffling on licenses.

Cohere’s approach — Apache 2.0, free, straight up “built for agents” — is sincere. Whether the North family will release bigger Code models or a Code-Plus to compete with Claude Sonnet is worth watching.

At least today, they’ve lowered the barrier for “running a local coding agent.” For indie developers and small teams valuing data sovereignty, this is genuinely beneficial.

References

• linux.do community discussion: openrouter new free model cohere/north-mini-code:free — first-hand user feedback
• Hugging Face: CohereLabs/North-Mini-Code-1.0 — official weights download and model card