North Mini Code
Neural Network
North Mini Code is a Cohere agentic model for programming: sparse MoE architecture with 30B parameters and 3B active.
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 256 000 tokens
- Max answer length
- 64 000 tokens
- Reasoning effort
- low, medium, high
- Capabilities
- reasoning, function calling
How it works North Mini Code?
North Mini Code is the first agentic model for code from Cohere and the debut model of the North family. Under the hood is a sparse mixture-of-experts architecture: out of 30 billion parameters, only 3 billion are active at each step, making the model designed for agentic development scenarios where multi-call loop performance is critical. In BotHub, North Mini Code is available from Russia without a VPN or foreign card; payment is made with Russian cards based on usage, and unused tokens do not expire. Nearby in the same window are over 250 other neural networks, so you can compare responses and switch to another model via a unified OpenAI-compatible API without rewriting the integration. Dialogues are encrypted via AES-GCM, and for companies, we offer contracts, invoices, EDI, and an admin panel with team limits. Connect the model as an agent in your code editor to let it read project files and make edits; assign it to refactor a legacy module or analyze confusing code; use it for debugging and stack trace analysis; or build an internal developer assistant using the BotHub API.Frequently asked questions about North Mini Code
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
This is a Cohere text model focused on code: writing functions, refactoring modules, finding bugs, explaining third-party repositories, or compiling documentation. The 256,000-token context window allows you to load a large project at once. It also accepts images and documents as input.
Up to 64,000 tokens in a single response — enough for a bulky code module, detailed technical analysis, or a large document. If the task is larger, break it into parts: the 256,000-token context allows you to keep the entire previous result in the same dialogue.
Yes, reasoning before answering is confirmed by our data. The model first breaks down the condition step-by-step and only then provides the result. This is especially noticeable in algorithms, debugging, and multi-step tasks where it is important to derive the answer sequentially rather than guessing.
Yes, both function calling and structured JSON output are supported. The model is convenient as a foundation for agents that trigger external tools, databases, and services. In BotHub, it is available via a unified API, so you won't have to rewrite your integration.
Images and documents — yes: send a screenshot with an error, an architecture diagram, or a specification file and get a text analysis. Audio and video processing is not noted in our data, so it is more convenient to choose a specialized model for such files in the BotHub catalog.
We do not have specific data on languages, and we will not make things up. The most honest way is to run your typical query and compare the result with another model in the same window. In BotHub, it takes a couple of clicks, without a VPN or foreign card.