Modelpublic

FreedomAISVR/Qwen3-Coder-30B-A3B-MXFP4-MOE-GGUF

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

1likes453downloads

Model Card

Qwen3-Coder-30B-A3B-Instruct-GGUF

MXFP4_MOE Quantization

This repository contains MXFP4_MOE quantized GGUF of Qwen3-Coder-30B-A3B-Instruct.

Model Details

Property	Value
Base Model	Qwen3-Coder-30B-A3B-Instruct
Architecture	Qwen3MoE
Parameters	30B (3.6B active, 128 experts, 8 activated)
Quantization	MXFP4_MOE (OCP MXFP4 E2M1, block 32, shared 8-bit block exponents)
BPW	4.47
File Size	~17.1 GB
Context Length	32K

Download

bash

huggingface-cli download FreedomAISVR/Qwen3-Coder-30B-A3B-MXFP4-MOE-GGUF qwen3-coder-30b-a3b-mxfp4_moe.gguf --local-dir . --local-dir-use-symlinks False

Quantization Information

This model uses MXFP4 (Microscaling FP4) quantization via llama.cpp's MXFP4_MOE type:

—E2M1 format: 1 sign bit, 2 exponent bits, 1 mantissa bit
—Block size: 32 elements sharing an 8-bit block exponent
—Expert weights: Quantized to MXFP4 (3 ffn_exps tensors per layer)
—Attention weights: Quantized to Q8_0 (8-bit block quantization)
—Other weights: Kept in F32/F16

Verification

After download, verify the file:

bash

echo "9f5a07e402df2aa16b9b4fcee22b5132 *qwen3-coder-30b-a3b-mxfp4_moe.gguf" | md5sum -c

Credits

—Qwen Team for the base model
—llama.cpp for the GGUF format and quantization tools
—MXFP4 is an OCP standard microscaling FP4 format backed by AMD, NVIDIA, Microsoft, Meta, and OpenAI