Cohere Releases North Mini Code 1.0: 30B-Parameter MoE Model With 256K Context for Agentic Coding

Cohere Labs has released North Mini Code 1.0, a 30B-parameter sparse Mixture-of-Experts (MoE) model with 3B active parameters designed for code generation and agentic software engineering tasks.

Model Architecture and Specifications

North Mini Code 1.0 uses a decoder-only Transformer architecture with 128 experts, activating 8 per token. The model features:

• Total parameters: 30B (3B active)
• Context window: 256K tokens with 64K max output
• License: Apache 2.0
• Architecture: Sparse MoE with interleaved attention (3:1 ratio of sliding-window with RoPE to global attention without positional embeddings)
• Training: Two-stage post-training with supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR)

Performance on Agentic Coding Benchmarks

Cohere evaluated the model on SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench v2, and Terminal-Bench Hard using the Swe-Agent harness v1.1.0. The company also tested on SciCode and LiveCodeBench v6 for complex code generation. All benchmarks used temperature=1.0 and top_p=0.95 across 3 seeds. Specific benchmark scores were not disclosed in the model card.

Tool Use and Integration

The model supports native tool-use capabilities through chat templates in Transformers. According to Cohere, North Mini Code 1.0 features "interleaved thinking" where the model generates reasoning content alongside tool calls. The company recommends passing all model-generated thinking content to future agentic steps for optimal performance.

Integration requires installing Transformers from source and, for vLLM deployment, using the main branch with Cohere's melody library (version 0.9.0+). The model uses a bash tool for terminal command execution.

Availability

North Mini Code 1.0 is available on Hugging Face and can be tested in OpenCode and Cohere's hosted Hugging Face Space. The model requires tensor parallelism (recommended -tp 2) for vLLM serving with a max model length of 320,000 tokens.

What This Means

This release targets the growing market for AI-powered software engineering tools, competing with models like GitHub Copilot and Amazon CodeWhisperer. The 30B-parameter count with 3B active parameters via sparse MoE suggests Cohere is prioritizing inference efficiency over raw model size—a practical choice for deployment in development environments. The Apache 2.0 license and open weights make this accessible for commercial use, though the lack of disclosed pricing and specific benchmark comparisons to competing models leaves performance questions open. The emphasis on agentic capabilities and tool use reflects the industry shift from simple code completion to multi-step reasoning and execution workflows.