This repository contains a backup of the Qwen 3.5-35B-A3B model quantized using Unsloth MXFP4 quantization. This is a backup/archive of a specific quantization version.

Overview

This is an unsloth MXFP4 quantized version of the Qwen 3.5-35B-A3B model. This repository serves as a backup/archive of this specific quantization.

Why This Exists

The primary/updated version of this model has been updated elsewhere. This repository is intentionally maintained as a stable backup copy of this specific unsloth MXFP4 quantization.

Purpose:

Preserve the original unsloth MXFP4 quantization without modifications
Archival copy for reference and reproducibility
Fallback option if the primary version changes or becomes unavailable
Stable checkpoint for experiments that require this exact version

Model Details

Base Model: Qwen 3.5-35B-A3B
Quantization: MXFP4 (via Unsloth)
License: Apache 2.0

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ayylmaonade/Qwen-3.5-35B-A3B-UD-MXFP4"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Vision Encoder / Multimodal Files

Important: When using this model for vision tasks, always prefer mmproj-F16.gguf over mmproj-F16-og-qwen3.5.gguf. The standard mmproj-F16.gguf was quantized with a newer version and provides superior performance compared to the older "OG" vision encoder.

Important Notes

⚠️ This is a backup version - if you need the most recent version, please check for the primary/updated repository.

This archive is maintained as-is to preserve the exact quantization without any modifications or updates.

References

Base Model: Qwen 3.5 from Alibaba Cloud
Quantization: Unsloth (https://github.com/unslothai/unsloth)
MXFP4: Memory-efficient 4-bit quantization format

License

Apache 2.0

ayylmaonade/Qwen3.5-35B-A3B-UD-MXFP4-GGUF

Qwen 3.5-35B-A3B (Unsloth MXFP4 Quantization) - Backup Archive