North¶

Development line: project:cohere-north · thread north-mini-code
Last event: 2026-06-30 · 1 dated since 2026-06-30 · Researched: 2026-09-05 · confidence: high

What it is¶

Cohere North Mini Code is a text sparse-MoE model for engineers who need a local alternative to small agentic code models. - Code generation. - Terminal tool execution. - Multi-step agentic tasks. - Structured tool calls.

30B parameters total, 3B active, 256K context, generation up to 64K tokens. It fits controlled coding-agent harnesses, but requires a compatible runtime and correct propagation of tool and thinking state.

Development line¶

• 2026-06-30 — CohereLabs published North Mini Code 1.0 model pages on Hugging Face. On 2026-06-30, the development line points to Hugging Face pages for North-Mini-Code-1.0 and a w4a16 variant. This marks a material distribution step for the North Mini Code line because it identifies both a base model artifact and a separately named variant.

What changed¶

• 2026-06-09 — Cohere released North-Mini-Code-1.0 under Apache 2.0 as the first North family model for agentic workflows, providing BF16, FP8, and W4A16 weights.
• 2026-06-30 — Repositories were recorded for the base North-Mini-Code-1.0 model and its W4A16 variant, confirming them as formats of one release rather than two separate models.

How to use this¶

From 2026-06-30, check the distinct Hugging Face pages for the base North-Mini-Code-1.0 artifact and its w4a16 variant rather than assuming a single undifferentiated model package.

1. Load base weights via Transformers, apply the built-in chat template to messages, and generate responses using sampling parameters from the model card. — https://huggingface.co/CohereLabs/North-Mini-Code-1.0
2. Run the model through vLLM and call /v1/chat/completions for an OpenAI-compatible local endpoint. — https://huggingface.co/CohereLabs/North-Mini-Code-1.0
3. Define functions with JSON Schema for a tool-using agent, pass tool results with the tool role, and continue generation. — https://huggingface.co/CohereLabs/North-Mini-Code-1.0

Best practices¶

• Set temperature=1.0 and top_p=0.95 as recommended by Cohere for generation. — https://huggingface.co/CohereLabs/North-Mini-Code-1.0
• Pass interleaved-thinking content into subsequent agent steps and conversation turns, as the model card notes this improves output quality. — https://huggingface.co/CohereLabs/North-Mini-Code-1.0
• Check version compatibility for vLLM: the model card specifies vLLM main and cohere_melody>=0.9.0 so responses parse correctly. — https://huggingface.co/CohereLabs/North-Mini-Code-1.0

Superseded by this¶

• 2026-06-30 — Treating North-Mini-Code-1.0 and North-Mini-Code-1.0-w4a16 as independent releases is obsolete: W4A16 is a weight format for North Mini Code 1.0.

Still unknown¶

• Primary sources confirm the 2026-06-09 release, but do not explain what distinct change occurred on 2026-06-30; that date may reflect repo recording or link publication rather than a new model release.
• The system response schema lacks event_findings and new_events fields; details on earlier releases and event clarification appear in what_changed and unknowns.

Sources¶

Agent brief¶

• Subject: project:cohere-north, thread north-mini-code, 1 dated events 2026-06-30 → 2026-06-30.
• Practical note: From 2026-06-30, practitioners assessing North Mini Code should check the distinct Hugging Face pages for the base North-Mini-Code-1.0 artifact and its w4a16 variant, rather than assuming a single undifferentiated model package.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.