North-Mini-Code-1.0 Materializes As An Open Agentic Coding Engine
North-Mini-Code-1.0 is a new open-source AI model from CohereLabs built specifically for code generation, autonomous software engineering, and terminal tasks. It packs a 30-billion-parameter architecture but only activates 3 billion parameters at a time, keeping it efficient. The model can handle up to 256,000 tokens of input and produce up to 64,000 tokens in a single response.
Cohere developed this release and shared it under the Apache 2.0 license. Their training focused on agentic coding, where the model learns to use tools like a terminal emulator to fix bugs or write programs on its own. The process combined supervised fine-tuning with reinforcement learning that rewards verified correct actions.
Efficient code agent with tool use
Key Features
- 30 billion total, 3 billion active parameters.
- Processes up to 256K tokens of input.
- Handles agentic coding and terminal tasks.
- Mixture-of-Experts with 128 total experts.
- Native tool calling via JSON schema definitions.
- Reinforcement learning with verifiable rewards.
- Open weights under Apache 2.0 license.
- Compatible with Hugging Face Transformers and vLLM.
This model suits developers and researchers building autonomous coding assistants or command-line tools. It can power applications that need to understand large codebases, write scripts, or interact with terminals. Since it’s open weights, teams can fine-tune it for specialized programming workflows.
Model performance and training insights
The model was evaluated on benchmarks like SWE-Bench Verified, Terminal-Bench, and LiveCodeBench, showing strong results on agentic coding tasks. It uses interleaved sliding-window and global attention, and its expert routing applies a sigmoid function before selecting the top experts. The team recommends using a temperature of 1.0 and top_p of 0.95 during generation, and passing the model’s reasoning content to future steps for best performance.
"North-Mini-Code-1.0 was post-trained using a two-stage cascaded supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR), focusing on agentic coding." — Source: Hugging Face
Project Links
