AB
AB
AiBoss

All Categories

project

North Mini Code - Cohere's open-source Agentic coding model

North Mini Code is an open-source agentic coding model from Cohere. The model uses the MoE architecture, has a total of 30 bytes of parameters, only 3 bytes of activation, and supports 256K context and 64K output.

What is the North Mini Code?

North Mini Code is an open-source agentic coding model from Cohere. The model uses the MoE architecture, has a total of 30 bytes of parameters and only 3 bytes of activation, supporting 256K context and 64K output. Optimized for code generation and software engineering tasks, it scored 33.4 on the Artificial Analysis Coding Index, with an API output speed of approximately 199 tokens/second, and can be deployed locally on a single H100 form.

Main functions of North Mini Code

  • Code generationThe model supports complex code generation tasks for scientific computing and algorithmic reasoning.
  • Agentic Software EngineeringIt can perform repository-level code modifications within frameworks such as SWE-Agent and OpenCode.
  • Terminal AgentEnd-to-end driven Shell tools to complete multiple rounds of build and deployment tasks.
  • Local encodingThe 3B activation parameter supports low-latency local inference, making it suitable for privacy-sensitive scenarios.
  • Sub-proxy orchestrationUnderstand the system architecture and coordinate with sub-agents to complete code review and refactoring.

The technical principles of North Mini Code

  • MoE Sparse ActivationModel 30B's total parameters only activate route 3B, significantly reducing inference cost and GPU memory usage.
  • Extremely long contextThe 256K input context supports loading a medium-sized codebase in one go, and the 64K output can generate a complete module.
  • Agentic trainingIt is designed for training on multiple frameworks such as SWE-Agent and ReAct terminal toolchain, and generalizes to different Agent scaffolding.
  • Speed optimizationUnder the same concurrency and hardware conditions, the output throughput is 2.8 times higher than Devstral Small 2, and the inter-token latency is 30% lower.

How to use North Mini Code

  • Hugging Face download weight: Obtain open-source weights and deploy locally using a compatible framework.
  • Cohere API CallsUse the Chat V2 API for a free trial with zero cost within your free quota.
  • Model Vault HostingEnterprise users can deploy production environments with a single click on the Cohere managed inference platform.
  • OpenCode Integration: Select directly on the OpenCode platform north-mini-code-free Try the model for free.

North Mini Code's core advantages

  • High efficiency and low costThe activation parameters of Model 3B make the inference cost much lower than that of similarly high-performance dense models.
  • Enterprise compliance friendlySupports on-premises, VPC, or private cloud deployments to meet the data ex-domain requirements of industries such as finance.
  • Advantages of a long windowThe 256K/64K context and output lengths are suitable for large codebase analysis and batch generation.
  • Multi-framework compatibilityNatively compatible with OpenCode, and also compatible with mainstream Coding Agent scaffolding.

North Mini Code's project address

  • Project official website: https://docs.cohere.com/docs/north-mini-code-1.0
  • HuggingFace model libraryhttps://huggingface.co/blog/CohereLabs/introducing-north-mini-code

North Mini Code's Competitive Comparison

Dimension North Mini Code Devstral Small 2
Architecture MoE (30B / 3B) Dense Model
Output throughput 2.8 times higher benchmark
Inter-Token Delay 30% lower benchmark
Maximum Context 256K shorter
Maximum output 64K shorter
Open source license Apache 2.0 Depending on the version
position Agentic code + terminal task General code assistance

Application scenarios of North Mini Code

  • Enterprise Private Code AssistantThe model supports local or VPC deployment, meeting the compliance requirements of industries such as finance and healthcare that core code should not leave the domain.
  • Repository-level refactoringThe 256K context window can load a medium-sized codebase at once, performing cross-file dependency analysis and batch refactoring.
  • Terminal automated production line: Drive the Shell toolchain through frameworks such as SWE-Agent to complete automated build, testing and deployment tasks.
  • Scientific computing assistanceIt generates algorithm implementations, data processing scripts, and complex mathematical reasoning code for researchers.
  • IDE real-time pluginThe 3B activation parameter supports low-latency inference, making it suitable for embedding in editors to provide real-time code completion and intelligent suggestions.
Back