- input
- $0.00
- output
- $0.00
- Context
- 256Ktokens
- Listed
- 4mo ago
- Size
- 30.5Bparameters
Cost/1M tokens
North Mini Code is an open-weights model from Cohere and the first release in the company's North family of code agent models. Cohere describes it as its first agentic coding model and says internal testing showed up to 2.8x higher output throughput than Devstral Small 2. It is built for code generation, agentic software engineering, and terminal tasks.
- Language model
- Open weights · 30.5B
- Reasoning
- Tools
How it compares
Additional source evidence
- Source model
- cohere/north-mini-code-1-0
- Source family
- north
- Last updated
- 2026-06-09
- Knowledge cutoff
- 2025-09-23
- Model context
- 256K
- Model output limit
- 64K
- Open weights
- Yes
- File attachments
- No
- Reasoning
- Yes
- Tool calling
- Yes
- Temperature control
- Yes
Published benchmarks
| Benchmark | Score | Evaluation details | Evidence |
|---|---|---|---|
| SWE-Bench Verified | 67.6 · resolved | SWE-agent · 2026-06-09 | Source |
| SWE-Bench Pro | 40.2 · resolve rate | SWE-agent · 2026-06-09 | Source |
| Artificial Analysis Intelligence Index | 27.6 · index score | 2026-06-09 | Source |
| Artificial Analysis Coding Index | 33.4 · index score | 2026-06-09 | Source |
| GDPval-AA | 14 · win rate | 2026-06-09 | Source |
| τ²-Bench Telecom | 37 · success rate | 2026-06-09 | Source |
Complete source record (JSON)
{
"id": "cohere/north-mini-code-1-0",
"name": "North Mini Code",
"description": "Cohere coding model for practical software engineering and agentic edits",
"family": "north",
"last_updated": "2026-06-09",
"knowledge": "2025-09-23",
"attachment": false,
"reasoning": true,
"tool_call": true,
"temperature": true,
"modalities": {
"input": [
"text"
],
"output": [
"text"
]
},
"limit": {
"context": 256000,
"output": 64000
},
"benchmarks": [
{
"name": "SWE-Bench Verified",
"score": 67.6,
"source": "https://huggingface.co/CohereLabs/North-Mini-Code-1.0",
"metric": "resolved",
"harness": "SWE-agent",
"date": "2026-06-09"
},
{
"name": "SWE-Bench Pro",
"score": 40.2,
"source": "https://huggingface.co/CohereLabs/North-Mini-Code-1.0",
"metric": "resolve rate",
"harness": "SWE-agent",
"date": "2026-06-09"
},
{
"name": "Artificial Analysis Intelligence Index",
"score": 27.6,
"source": "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model",
"metric": "index score",
"date": "2026-06-09"
},
{
"name": "Artificial Analysis Coding Index",
"score": 33.4,
"source": "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model",
"metric": "index score",
"date": "2026-06-09"
},
{
"name": "GDPval-AA",
"score": 14,
"source": "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model",
"metric": "win rate",
"date": "2026-06-09"
},
{
"name": "τ²-Bench Telecom",
"score": 37,
"source": "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model",
"metric": "success rate",
"date": "2026-06-09"
}
],
"release_date": "2026-06-09",
"open_weights": true
}Frequently asked questions
Running locally
About 22.3 GB at Q4, 37.3 GB at Q8 or 70.6 GB at FP16, all estimates from the Ollama build sizes. For the 30.5B parameter model these estimates include weights plus about 15% and 0.5 GB for a short context.
Yes, a 24 GB GPU fits at Q4 with an estimated 22.3 GB. A 16 GB, 12 GB or 8 GB GPU does not fit at Q4, Q8 or FP16 (estimate).
Yes, assuming macOS lets the GPU use about 75% of unified memory, the smallest common Mac memory size that fits the estimated 22.3 GB at Q4 is 32 GB.
Q4_K_M, in the Ollama build north-mini-code-1.0:latest. It is the smallest published build at 4-bit or higher, with a 19 GB download and an estimated 22.3 GB of VRAM.
Run ollama run north-mini-code-1.0:latest. That build is a 19 GB download, and there are 5 other local Ollama builds.
No, the only runtime with recorded support is Ollama. North Mini Code has published Ollama builds, but no recorded support for llama.cpp, vLLM, LM Studio or MLX.
The default Ollama build north-mini-code-1.0:latest is 19 GB. Across the 6 builds with a published size, the smallest is 19 GB and the largest is 61 GB.
Privacy and data
Yes, North Mini Code has open weights so you can self-host it, for example with ollama run north-mini-code-1.0:latest.
Capabilities and specs
The context window is 256K tokens (256,000) and the maximum output is 64K tokens (64,000).
No, the only input modality is text. North Mini Code cannot read images, audio, video or PDFs.
Yes, North Mini Code supports tool calling (function calling).
Yes, North Mini Code supports a reasoning mode, but reasoning is not always on.
The knowledge cutoff of North Mini Code is 2025-09-23.
Quality and benchmarks
Kimi K2.7 Code Highspeed by Moonshot AI, Ornith 1.0 31B by DeepReinforce and Ornith 1.0 35B by DeepReinforce. They are matched on model kind and release date, with no comparable price.
Kimi K2.7 Code HighspeedNorth Mini Code vs Kimi K2.7 Code HighspeedOrnith 1.0 31BNorth Mini Code vs Ornith 1.0 31BOrnith 1.0 35BNorth Mini Code vs Ornith 1.0 35B
Licensing and access
North Mini Code has open weights.
The weights are on Hugging Face in the repository CohereLabs/North-Mini-Code-1.0. There are also 6 Ollama builds.
Basics
Cohere made North Mini Code, which was first listed on OpenRouter on June 17, 2026.
North Mini Code is an open weights model.
Generated from Vertical Tabs data, updated November 3, 2026.
More in North
Discussion
No comments yet
Be the first to share your thoughts.