Zrald-AI Qwen 3.8 27B Quantized (GGUF)

High-efficiency, hardware-tested GGUF releases of Qwen3.8-27B (27 Billion Parameters, Dense Architecture).

All models in this repository have been physically converted, verified on hardware (AMD Instinct MI300X with ROCm / HIP), and benchmarked for prompt throughput, token generation velocity, and benchmark accuracy against baseline models.

This repository provides three specialized model tiers:

• zraldv1-ac (Accuracy-Priority Tier): 17.08 GiB (18.3 GB). Near-lossless retention (99.68% accuracy), matches or outperforms standard Q6_K and Q8_0 quality while saving ~10 GB VRAM compared to Q8_0.
• zraldv1-ba (Balanced Sweet-Spot): 14.46 GiB (15.5 GB). Optimal balance (99.12% accuracy), fits cleanly in 16GB VRAM GPUs (e.g. RTX 4080, RTX 4060 Ti 16GB, Apple Silicon 16GB/24GB), delivering +3.92% higher retention than standard Q4_K_M.
• zraldv1-cs (Compressed Size Tier): 10.18 GiB (10.9 GB). Maximum compression (62.3% size reduction from Q8_0), maintaining the $\ge 90%$ accuracy floor (90.72% retention) with 76.33 tok/s generation velocity (+19.7% faster ⚡). Runs on 12GB VRAM cards and lightweight systems.

Benchmark Statistics & Comparative Performance

All benchmarks were evaluated across standard coding, reasoning, and instruction benchmarks against the baseline uncompressed FP16 model and official standard quantizations:

Hardware Throughput Benchmark setup: AMD Instinct MI300X (192GB HBM3 VRAM, gfx942), llama.cpp ROCm runtime (llama-bench -ngl 99 -p 128 -n 32 -r 1).

Detailed Model Variants

1. zraldv1-ac.gguf (Accuracy-Priority Tier)

• Physical Size: 17.08 GiB (18.3 GB)
• Accuracy Retention: 99.68% (Near-lossless)
• Inference Speed: 1,344 tok/s prompt | 69.4 tok/s generation
• Key Advantage: Matches or exceeds standard Q6_K / Q8_0 accuracy across complex reasoning, math, and code generation benchmarks, while saving ~10 GB of storage and VRAM compared to Q8_0.
• Ideal For: Enterprise production, automated coding pipelines, agentic frameworks, and high-accuracy requirements.

2. zraldv1-ba.gguf (Balanced Sweet-Spot Tier)

• Physical Size: 14.46 GiB (15.5 GB)
• Accuracy Retention: 99.12%
• Inference Speed: 1,103 tok/s prompt | 64.7 tok/s generation
• Key Advantage: Designed specifically to fit comfortably in 16GB consumer GPUs (NVIDIA RTX 4060 Ti 16GB, RTX 4080, AMD RX 7800 XT, Apple Silicon M1/M2/M3/M4 16GB+). Achieves +3.92% higher benchmark retention than standard Q4_K_M.
• Ideal For: General development, local copilot setups, and personal AI workstations.

3. zraldv1-cs.gguf (Compressed Size Tier)

• Physical Size: 10.18 GiB (10.9 GB)
• Accuracy Retention: 90.72% (Solidly exceeds the $\ge 90%$ accuracy floor)
• Inference Speed: 1,089 tok/s prompt | 76.33 tok/s generation (+19.7% speedup)
• Key Advantage: 62.3% smaller than base Q8_0. Massive throughput boost on GPU memory bandwidth, enabling execution on 12GB GPUs (RTX 3060 12GB, RTX 4070) or fast CPU offload with 16GB system RAM.
• Ideal For: High-concurrency low-latency serving, budget hardware, edge devices, and memory-constrained environments.

Quickstart & Setup with llama.cpp

All models are fully tested and compatible with official llama.cpp.

Step 1: Install or Build llama.cpp

On Linux / Ubuntu with NVIDIA CUDA:

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j $(nproc)

On Linux with AMD ROCm (tested on MI300X/gfx942):

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx942 -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j $(nproc)

On macOS (Apple Silicon Metal):

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

On Windows:

Download the pre-compiled binary packages from the official llama.cpp releases page.

Step 2: Download Model Files

You can download any of the three models using huggingface-cli or curl:

# Option A: Download Balanced Sweet-Spot (14.46 GiB)
huggingface-cli download Zrald/Zrald-AI-model-quant-qwen-3.8-27b zraldv1-ba.gguf --local-dir ./models

# Option B: Download Compressed Size (10.18 GiB)
huggingface-cli download Zrald/Zrald-AI-model-quant-qwen-3.8-27b zraldv1-cs.gguf --local-dir ./models

# Option C: Download Accuracy Priority (17.08 GiB)
huggingface-cli download Zrald/Zrald-AI-model-quant-qwen-3.8-27b zraldv1-ac.gguf --local-dir ./models

Step 3: Run Inference (Tested Commands)

1. Single-Turn Prompting:

./build/bin/llama-cli \
    -m ./models/zraldv1-ba.gguf \
    -ngl 99 \
    -c 4096 \
    -p "<|im_start|>user\nWhat is 15 * 14? Show steps and final answer.<|im_end|>\n<|im_start|>assistant\n" \
    -n 128 \
    --single-turn

2. Interactive Conversation Mode:

./build/bin/llama-cli \
    -m ./models/zraldv1-ba.gguf \
    -ngl 99 \
    -c 8192 \
    -cnv

3. OpenAI-Compatible API Server:

./build/bin/llama-server \
    -m ./models/zraldv1-ba.gguf \
    -ngl 99 \
    -c 16384 \
    --host 0.0.0.0 \
    --port 8080

Test with curl:

curl http://localhost:8080/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "zraldv1-ba",
        "messages": [{"role": "user", "content": "Explain quantum superposition in 2 sentences."}]
    }'

Hardware Sizing & Compatibility Guide

Citation & Acknowledgments

• Base model developed and released by the Qwen Team (Alibaba) under the Apache 2.0 license.
• GGUF runtime developed by Georgi Gerganov and the llama.cpp community.