Original Model: XYZAILab/XYZ-Aquila-mini
Base Architecture: Qwen3.6-35B-A3B (MoE, ~3B active parameters per token)
Model Type: Open-weight thinking model optimized for Deep Search and Agentic Workflows.
Quantization Formats: GGUF (IQ4_XS, IQ4_NL, Q4_K_M, Q4_K_S)

Quantization Details

This repository specifically focuses on 4-bit quantizations to maximize the balance between VRAM footprint and reasoning capability.

File Name	Size	Quant Method	Description
`XYZ-Aquila-mini-IQ4_XS.gguf`	~18.8 GB	`IQ4_XS`	4-bit importance matrix (extra small). Optimized for low VRAM while retaining reasoning quality.
`XYZ-Aquila-mini-IQ4_NL.gguf`	~19.9 GB	`IQ4_NL`	4-bit importance matrix (non-linear). Improved non-linear weight distribution for accuracy.
`XYZ-Aquila-mini-Q4_K_M.gguf`	~21.4 GB	`Q4_K_M`	Standard 4-bit K-quantization (medium). Recommended general-purpose balance of speed and quality.
`XYZ-Aquila-mini-Q4_K_S.gguf`	~20.6 GB	`Q4_K_S`	Standard 4-bit K-quantization (small). Slightly smaller file size.

Usage Guide

XYZ-Aquila-mini is a thinking model with Qwen-compatible reasoning and tool-call formats. It performs best when allowed to output chain-of-thought reasoning before acting.

1. Running with `llama.cpp`

You can run this model natively using llama-server to expose an OpenAI-compatible API that supports tool calling.

bash

# Clone and build llama.cpp
git clone [https://github.com/ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp)
cd llama.cpp && cmake -B build && cmake --build build -j

# Download the recommended model from this repository
huggingface-cli download Abiray/XYZ-Aquila-mini-IQ4-GGUF XYZ-Aquila-mini-Q4_K_M.gguf --local-dir .

# Serve the model (adjust context length `-c` based on your available RAM/VRAM)
./build/bin/llama-server \
  -m XYZ-Aquila-mini-Q4_K_M.gguf \
  --port 8000 \
  -c 32768 \
  -ngl 99

Abiray/XYZ-Aquila-mini-IQ4-GGUF

XYZ-Aquila-mini - 4-bit GGUF Quants

Quantization Details

Usage Guide

1. Running with llama.cpp

Abiray/
XYZ-Aquila-mini-IQ4-GGUF

1. Running with `llama.cpp`