Description
North Mini Code is a sophisticated model developed by Cohere and Cohere Labs, optimized for code generation, agentic software engineering, and terminal tasks. It features a 30 billion parameter model with 3 billion active parameters, making it highly efficient for complex coding tasks. The model supports a context length of 256K and a maximum output length of 64K, providing ample capacity for extensive coding projects.
Read More
The model architecture is based on a decoder-only Transformer with a sparse Mixture-of-Experts design. It incorporates an efficient attention mechanism, interleaving sliding-window attention with RoPE and global attention without positional embeddings. This design is complemented by a feed-forward block with 128 experts, of which 8 are activated per token, ensuring precise and effective processing.
North Mini Code has been post-trained using a two-stage cascaded supervised fine-tuning followed by reinforcement learning with verifiable rewards, focusing on agentic coding. This training approach enhances its ability to handle complex coding tasks and tool-use capabilities. The model is open-source, licensed under Apache 2.0, and available for use through Hugging Face's platform.
Users can try North Mini Code in OpenCode and the hosted Hugging Face Space, allowing for hands-on experience before downloading the model weights. The model's performance has been benchmarked using various methodologies, including SWE-Bench and Terminal-Bench, demonstrating its proficiency in agentic coding tasks.
North Mini Code is ideal for developers and engineers seeking a robust model for code generation and software engineering tasks. Its tool-use capabilities are supported through chat templates in Transformers, enabling seamless integration into existing workflows.
- CategoryAI Coding Assistants
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
North Mini Code Highlights
30B-A3B parameter model
Optimized for code generation
Supports agentic software engineering
Sparse Mixture-of-Experts architecture
Efficient attention implementation
256K context length
64K max output length
Open-source under Apache 2.0
Getting Started with North Mini Code
Access model page: Visit Hugging Face
Authenticate: Log in to your account
Configure: Set sampling parameters
Send prompts: Use transformers pipeline
Fine-tune: Apply supervised fine-tuning
North Mini Code's Use Cases
- Code Generation
- Software Engineering
- Terminal Tasks
- Tool Integration
- Benchmark Testing
