North Mini Code
Cohere's first open-weight agentic coding model - 30B MoE, 3B active, runs on one H100.
Context window
256K tokens
Input / 1M tokens
0
Output / 1M tokens
0
Provider
Cohere
Open-weight under Apache 2.0 - free to self-host (note the Hugging Face card adds a non-commercial usage note). Also available via the Cohere API, Cohere Model Vault (managed inference), and OpenRouter, which set their own rates. · Data verified 2026-07-02
North Mini Code is Cohere's first open-weight agentic coding model, released June 2026. It is a 30B-total / 3B-active mixture-of-experts model under Apache 2.0, with a 256K-token context window and 64K max generation length. Built and trained against multiple agent harnesses (SWE-agent, ReAct terminal agents), it generalizes across frameworks rather than overfitting one, and delivers up to 2.8x higher output throughput than Devstral Small 2 on the same hardware. Its small active footprint lets it run on a single H100 for local, sovereign deployment.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Download open weights from Hugging Face (CohereLabs/North-Mini-Code-1.0, in bf16/fp8/w4a16), run locally via Ollama, or call it through the Cohere API, Cohere Model Vault, or OpenRouter. Compact enough to run on a single H100.
Strengths
- ✓Open weights (Apache 2.0) with a tiny 3B active footprint - runs on a single H100
- ✓Purpose-built for agentic software engineering and terminal-based agents
- ✓Trained across multiple harnesses, so performance generalizes between agent frameworks
- ✓256K context window for repo-level understanding
- ✓Up to 2.8x higher throughput and ~30% lower inter-token latency vs Devstral Small 2
Best for developers who...
When to choose it (and when not to)
Reach for North Mini Code when...
- →You want a self-hostable, sovereign coding agent on modest hardware
- →You run agentic SWE workflows (repo edits, sub-agent orchestration, code review)
- →You need terminal/tool-driven multi-turn agents
- →You want high coding throughput at low compute cost
Look elsewhere if...
- ✕You need the absolute top SWE-bench scores from a large frontier coding model
- ✕Your use is strictly commercial and the HF card's non-commercial note is a concern
- ✕You need general-purpose chat or multimodal input rather than coding
- ✕You need a fully managed model with enterprise SLAs out of the box
How to use it
- ›Run it inside an agent harness (e.g. SWE-agent or a ReAct terminal loop) to get its intended behavior
- ›Give it repo context and clear task specs for repo-level changes
- ›Expose shell/file tools for terminal-based multi-turn workflows
- ›Use fp8 or w4a16 weights to fit comfortably on a single H100
Quickstart
Pythonimport cohere
co = cohere.ClientV2(api_key="YOUR_COHERE_API_KEY")
resp = co.chat(
model="North-Mini-Code-1.0",
messages=[{"role": "user", "content": "Fix the failing test in utils/parser.py and explain the change."}],
)
print(resp.message.content[0].text)Also fully self-hostable: pull CohereLabs/North-Mini-Code-1.0 from Hugging Face (or Ollama) and serve locally on a single H100.
API model id: north-mini-code-1-0
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Artificial Analysis Coding Index | 33.4 | Reported by Cohere; competitive among similarly sized coding models. |
| Throughput vs Devstral Small 2 | Up to 2.8x higher output throughput | Under identical concurrency and hardware; also ~30% edge in inter-token latency. |