vincespeed/Ling-3.0-tiny-APEX-GGUF

Ling-3.0 Tiny — Apex Quant GGUF Models

This repository contains 4 quantized GGUF profiles of the inclusionAI/Ling-3.0-tiny model, produced using Apex-Quant technology.

📦 Model Profile Summary

Profile	Size	BPW	Use Case
i-quality	5.4 GB	5.83	Highest quality, production environments
i-balanced	5.6 GB	6.02	Balanced quality and performance
i-compact	3.7 GB	4.03	Compact deployment, low RAM

BPW = Bits Per Weight. Higher value = better quality.

📁 File Structure

models/
├── Ling-3.0-tiny-i-quality.gguf   # 5.4 GB — Highest quality
├── Ling-3.0-tiny-i-balanced.gguf  # 5.6 GB — Balanced
└── Ling-3.0-tiny-i-compact.gguf   # 3.7 GB — Compact

🔗 Source Model

These models were created based on the inclusionAI/Ling-3.0-tiny model from HuggingFace.

• Model Page: https://huggingface.co/inclusionAI/Ling-3.0-tiny
• Architecture: BailingMoeV3ForCausalLM (Mixture-of-Experts)
• Parameter Count: 128×1.0B (128 experts, each with 1B parameters)
• Context Length: 131,072 tokens
• Vocabulary: 157,184 tokens
• License: MIT

🛠️ Technology

These quantized models were produced using Apex-Quant technology.

• Apex-Quant: MoE-aware mixed-precision quantization
• Infrastructure: llama.cpp (llama-quantize)
• Quantize Script: apex-quant/scripts/quantize.sh

🙏 Acknowledgments

• localai-org/apex-quant — Apex-Quant MoE-aware mixed-precision quantization framework
• ggerganov/llama.cpp — GGUF format and quantization engine
• inclusionAI — Original Ling-3.0-tiny model creators

📋 Technical Details

Architecture Information

• Architecture: bailingmoe3
• Block Count: 24 layers
• Expert Count: 128 experts
• Expert Used Count: 8 experts/token
• Expert Group Count: 8
• Expert Group Used Count: 4
• Expert Gating Function: Top-K (k=8)
• Hidden Size: 1,536
• Feed Forward Size: 4,608
• Attention Heads: 16
• Attention Head Count KV: [0, 0, 0, 1, ...] (grouped query attention)
• Rope Frequency Base: 6,000,000
• Layer Norm Epsilon: 1e-6

Quantize Profile Details

i-quality (Q6_K/Q5_K/IQ4_XS)

• Expert FFN: Q6_K / Q5_K / IQ4_XS (mixed)
• Shared FFN: Q8_0
• Attention: Q6_K
• BPW: 5.83
• File Size: 5.4 GB

i-balanced (Q6_K/Q5_K)

• Expert FFN: Q6_K / Q5_K (mixed)
• Shared FFN: Q8_0
• Attention: Q6_K
• BPW: 6.02
• File Size: 5.6 GB

i-compact (Q4_K/Q3_K)

• Expert FFN: Q4_K / Q3_K (mixed)
• Shared FFN: Q6_K
• Attention: Q4_K
• BPW: 4.03
• File Size: 3.7 GB

💻 Usage

With llama.cpp

# Run with i-quality profile
./main -m models/Ling-3.0-tiny-i-quality.gguf -n 128 -p "Hello, how are you?"

# Run with i-compact profile
./main -m models/Ling-3.0-tiny-i-compact.gguf -n 128 -p "Hello, how are you?"

With Ollama

# Create Dockerfile or Ollamafile
FROM llama.cpp
COPY models/Ling-3.0-tiny-i-quality.gguf /model.gguf

With Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="models/Ling-3.0-tiny-i-quality.gguf",
    n_ctx=4096,
    n_threads=8
)

output = llm(
    "Hello, how are you?",
    max_tokens=128
)
print(output["choices"][0]["text"])

📊 Model Comparison

Criterion	i-quality	i-balanced	i-compact
Quality	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	⭐⭐⭐
Speed	⭐⭐⭐	⭐⭐⭐⭐	⭐⭐⭐⭐⭐
RAM	High	Medium	Low
Size	5.4 GB	5.6 GB	3.7 GB
BPW	5.83	6.02	4.03

📝 Notes

• All models are in GGUF v3 format.
• The BailingMoeV3 architecture uses Mixture-of-Experts (MoE) technology.
• The model uses grouped query attention (GQA) and rope positional embeddings.
• The i-mini profile cannot be quantized without imatrix. ~100-200 inference samples must be run on the model to generate the importance matrix.

📄 License

The original model is distributed under the MIT license. The quantized models are shared under the same license.

🔗 Related Links

• Original Model: https://huggingface.co/inclusionAI/Ling-3.0-tiny
• Apex-Quant: https://github.com/localai-org/apex-quant
• llama.cpp: https://github.com/ggerganov/llama.cpp
• GGUF Format: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md

---

Note: These models are quantized for local use. Check the original model's license for commercial use.