---

license: apache-2.0

language:

• en

library_name: gguf

tags:

• gguf
• cohere
• north
• code
• moe
• mxfp4
• open-standard

base_model: CohereLabs/North-Mini-Code-1.0

pipeline_tag: text-generation

inference: false

quantized_by: freedom11

---

North-Mini-Code-1.0-MXFP4-GGUF

MXFP4_MOE (OCP Microscaling FP4 E2M1, block size 32) 4-bit quantization of CohereLabs/North-Mini-Code-1.0.

Model Description

North-Mini-Code-1.0 is a 30B total parameter MoE code model with 2.7B active parameters per token. It uses 128 experts with 8 selected per token, 49 transformer layers (hybrid sliding window + full attention at 3:1 ratio), and a vocabulary of 256K tokens. Architecture follows the Cohere2MoE design with parallel residual blocks, grouped-query attention (32 heads, 4 KV heads, 8:1 GQA ratio), RMS norm, and SiLU-gated activations.

| Config | Value |

|--------|-------|

| Total parameters | ~30.5B |

| Active parameters | ~2.7B |

| Layers | 49 (13 full + 36 sliding window, 3:1 ratio) |

| Attention heads | 32 (4 KV heads, GQA 8:1) |

| Head dimension | 128 |

| Hidden dimension | 2048 |

| MLP intermediate | 768 (MoE), 3072 (dense prefix) |

| Experts | 128 (8 active per token) |

| Context window | 4096 (sliding) / 500000 (full with RoPE) |

| Vocabulary | 262144 tokens |

| RoPE theta | 50000.0 |

MXFP4_MOE Quantization

MXFP4_MOE applies the OCP MXFP4 microscaling format (E2M1, block size 32) to expert weight tensors while keeping attention projections at Q8_0. This hybrid approach optimizes the quality-size tradeoff for MoE architectures: the 128 expert FFN layers (~97% of parameters) benefit from MXFP4 density, while attention tensors stay higher precision for better routing and context processing.

Unlike NVFP4 (NVIDIA-proprietary), MXFP4 is an open OCP standard compatible with any GPU or CPU backend that implements the microscaling specification.

| Format | File Size | BPW | Block Size | Expert Format | Attention Format |

|--------|-----------|-----|------------|---------------|-----------------|

| MXFP4_MOE | 17.04 GB | ~4.8 | 32 | MXFP4 (E2M1) | Q8_0 |

| BF16 (original) | 56.8 GB | 16 | 1 | BF16 | BF16 |

Notes:

• MXFP4 per-block shared exponents preserve dynamic range for expert weight outliers
• Q8_0 attention layers maintain precision for key/value projection and output
• Compatible with any GPU supporting MXFP4 via CUDA 13.x or LLVM SPIR-V; falls back to CPU for unsupported hardware
• Open standard (OCP Microscaling, OCP Specification v1.0)

Files

| File | Size | Description |

|------|------|-------------|

| north-mini-code-1.0-mxfp4_moe.gguf | 17.04 GB | MXFP4_MOE quantized text model |

Conversion Pipeline

CohereLabs/North-Mini-Code-1.0 (HF safetensors, BF16, 56.8 GB)
  -> convert_hf_to_gguf.py --outtype f16 (GGUF F16, 61.0 GB, cohere2_moe arch)
  -> llama-quantize.exe MXFP4_MOE (GGUF MXFP4_MOE, 17.04 GB, 442 tensors)

Usage

llama.cpp:

./llama-cli -m north-mini-code-1.0-mxfp4_moe.gguf -p "Write a Python function implementing merge sort with type annotations" -n 512 -t 8 -c 8192

llama-cpp-python:

from llama_cpp import Llama
llm = Llama(model_path="north-mini-code-1.0-mxfp4_moe.gguf", n_ctx=8192, n_threads=8, n_gpu_layers=-1)
output = llm("Write a Python function implementing merge sort with type annotations", max_tokens=512)
print(output["choices"][0]["text"])

Hugging Face Hub:

from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="FreedomAISVR/North-Mini-Code-1.0-MXFP4-GGUF", filename="north-mini-code-1.0-mxfp4_moe.gguf")

Hardware

Quantized on NVIDIA GeForce RTX 5060 Ti (16 GB VRAM, Blackwell). Conversion time: ~13 minutes.

License

Apache-2.0 (same as original model).