For engineering teams looking to build agentic coding pipelines, Cohere has released a significant open-source alternative: North Mini Code. This new model is designed specifically for software engineering tasks, moving beyond general-purpose models. It boasts impressive capabilities, including advanced tool-use, interleaved thinking for complex multi-step operations, and the ability to analyse and map system architectures, review code across vast codebases, and handle terminal-based agentic tasks.

A key feature of North Mini Code is its large context window of 256,000 tokens, allowing it to process substantial multi-file projects in a single go. This makes it particularly adept at tasks like architecture mapping and code review. The model has been trained on real-world terminal environments, as evidenced by its performance on the Terminal-Bench v2 benchmark, which tests agents in practical shell interactions rather than just synthetic code generation.
Technically, North Mini Code is a sparse mixture-of-experts (MoE) model. While it has a total of 30 billion parameters, only 3 billion are active per token, making its inference compute requirements comparable to a much smaller model. This efficiency allows it to run on a single H100 GPU, and even demonstrated on a Mac Studio. Cohere trained the model using a multi-stage process of supervised fine-tuning and reinforcement learning, focusing on verifiable rewards across a wide array of coding tasks and utilising multiple agent scaffolds for robustness. This approach aims to provide a more flexible and powerful agentic coding experience compared to proprietary, managed models.
Fuente Original: https://venturebeat.com/technology/cohere-open-sources-a-coding-agent-that-runs-on-a-single-h100
Artículos relacionados de LaRebelión:
- Europes Open Source
- VS Code Retrasa Actualizaciones Contra Ataques Ciberneticos
- EUs Open Source Revolution Tech Sovereignty Unveiled
- IBM and Red Hat Invest 5 Billion in Open Source Security
- Agentes IA en Peligro Vulnerabilidad Critica en Paquete Open Source
Artículo generado mediante LaRebelionBOT












