Cohere has released an open-source coding agent named North Mini Code, offering engineering teams a tangible alternative to managed models like Claude Fable 5. This model can run efficiently on a single H100, but it comes with a verbosity cost—producing three times the output tokens of similar models, which could be a concern for high-volume operations.
## What North Mini Code Can Do
North Mini Code is designed to cater to the entire agentic coding stack, offering a targeted approach to software engineering tasks. Unlike models adapted from general-purpose bases, North Mini Code is purpose-built for agentic software engineering, boasting integrated tool-use capabilities that enhance multi-step agentic work. This specificity could improve efficiency in complex coding environments.
The model shines in architecture mapping and code review, able to analyze systems, identify dependencies, and perform reviews across extensive codebases. Its 256,000 token context window allows it to handle large multi-file projects in a single pass, potentially reducing the time and effort required for comprehensive reviews.
Furthermore, North Mini Code is adept at handling terminal-based tasks, such as shell interactions and command-line tooling. Cohere validated its performance using Terminal-Bench v2, which assesses models in authentic terminal environments. This focus on real-world applications could make it a practical choice for engineers working in terminal-heavy settings.
## How It Was Built
North Mini Code is a sparse mixture-of-experts (MoE) model, featuring 128 experts with 8 active per token. Despite its 30 billion total parameters, the compute requirement during inference is akin to a 3 billion parameter model. This efficiency allows it to run on modest hardware, like a Mac Studio with 20 gigabytes of RAM.
The model underwent rigorous training, starting with two stages of supervised fine-tuning, followed by reinforcement learning across more than 70,000 tasks. These tasks spanned approximately 5,000 repositories, ensuring a wide-ranging and thorough training process. Cohere’s multi-harness training approach, which includes SWE-Agent, Mini-SWE-Agent, and OpenCode, reportedly led to a 10 percentage point improvement on OpenCode evaluations without sacrificing SWE-Agent performance.
## Where It Fits
North Mini Code enters a competitive market populated by models like Mistral Devstral Small 2, GitHub Copilot, Cursor, and Claude Fable 5. Each competitor presents unique cost and deployment considerations, making the landscape diverse and challenging to navigate.
Cohere positions North Mini Code against Mistral Devstral Small 2, a 24 billion parameter dense model. Cohere claims their model achieves 2.8 times higher output throughput and a 30% inter-token latency advantage over Devstral Small 2 in internal tests. These performance metrics suggest that North Mini Code could offer a viable alternative for teams seeking efficiency without sacrificing performance.
## What Happens Next
The release of North Mini Code provides a new option for engineering teams exploring open-source solutions for agentic coding tasks. For founders and engineers, the decision to adopt this model hinges on weighing verbosity costs against the potential efficiency gains in specific applications. Investors and VCs should consider the model’s potential to disrupt established players in a rapidly evolving market.
For those in the tech industry, especially those focused on software engineering and AI, North Mini Code’s open-source nature and specific capabilities may offer a new tool to enhance productivity. As the market continues to evolve, keeping an eye on performance benchmarks and real-world applications will be crucial for making informed decisions.
