---

language:

• en
• pl

license: apache-2.0

tags:

• polish
• chatgpt
• text-generation
• conversational
• llama.cpp
• gguf
• ollama
• lm-studio
• jan-ai
• qwen2
• 14b
• dus
• code

base_model: Szymon-Wachelko/GPT-neo-Custom-7B-GGUF

pipeline_tag: text-generation

inference: false

---

<div align="center">

⚡ GPT-neo-Custom (14B-DUS) GGUF ⚡

Next-Generation Polish & English AI Assistant (DUS Architecture)








<p align="center">

<b>Trained & Engineered with passion by Szymon-Wachelko</b><br>

<i>Built with Depth Up-Scaling (DUS 56 Layers) based on GPT-neo-Custom 7B and optimized for deep reasoning, programming, and native Polish eloquence.</i>

</p>

</div>

---

> [!TIP]

> 💬 Masz pomysł na ulepszenie modelu, chcesz coś dodać lub zgłosić błąd? / Found a bug or feature request?

> Wejdź w zakładkę Community / Discussions! Wszystkie zgłoszenia, propozycje ulepszeń oraz raporty są aktywnie monitorowane i wdrażane w kolejnych aktualizacjach.

---

📖 Model Overview

> GPT-neo-Custom 14B-DUS is a flagship high-performance conversational language model designed to bridge the gap in high-quality Polish language generation while maintaining top-tier software engineering and mathematical problem-solving skills.

Built upon the state-of-the-art Szymon-Wachelko/GPT-neo-Custom-7B-GGUF and scaled using the Depth Up-Scaling (DUS) technique to 56 neural layers (~14.7 billion parameters), this model delivers superior context comprehension, complex multi-turn dialogue memory, and abstract thinking.

✨ Key Strengths

• 🇵🇱 Natural Polish Fluency: Full mastery of Polish grammar, complex sentence structures, declensions, and cultural nuances without translation artifacts.
• 🧠 DUS Architecture (56 Layers): ~14.7B parameters provide deep contextual awareness, high logic retention, and robust multi-step problem solving.
• 💻 Software Engineering: Proficient in writing, debugging, and explaining code in Python, C++, JavaScript, TypeScript, SQL, Rust, and shell scripting.
• 🧮 Structured Reasoning: Excels at chain-of-thought mathematical reasoning, logic puzzles, and technical documentation.
• 🚀 Universal Compatibility: High-performance GGUF quants tailored for local execution from everyday laptops to dedicated RTX workstations.

---

📊 Quantization Matrix

| File Name | Precision | File Size | Recommended VRAM | Recommended Setup |

| :--- | :---: | :---: | :---: | :--- |

| GPT-neo-Custom-14B-F16.gguf | 16-bit Float | ~28.5 GB | 32 GB+ | Reference studio grade; 100% lossless precision. |

| GPT-neo-Custom-14B-Q8_0.gguf | 8-bit | ~15.2 GB | 18–20 GB | 🏆 Gold Standard: Extremely high accuracy; near-identical to FP16. |

| GPT-neo-Custom-14B-Q6_K.gguf | 6-bit | ~12.2 GB | 14–16 GB | High accuracy with blazing fast generation speeds on RTX 3090/4090. |

| GPT-neo-Custom-14B-Q5_K_M.gguf | 5-bit Medium | ~10.5 GB | 12–14 GB | Optimal balance between intelligence and resource usage. |

| GPT-neo-Custom-14B-Q5_K_S.gguf | 5-bit Small | ~9.9 GB | 12 GB | Compact 5-bit version for balanced performance. |

| GPT-neo-Custom-14B-Q4_K_M.gguf | 4-bit Medium | ~8.01 GB | 10–12 GB | 🌟 Most Popular / Recommended: Perfect for RTX 3060 (12GB) & consumer desktop inference. |

| GPT-neo-Custom-14B-Q4_K_S.gguf | 4-bit Small | ~7.5 GB | 10 GB | Lighter 4-bit variant for memory-constrained systems. |

| GPT-neo-Custom-14B-Q3_K_L.gguf | 3-bit Large | ~6.9 GB | 8–10 GB | Preserved attention heads for high quality at lower bitrates. |

| GPT-neo-Custom-14B-Q3_K_M.gguf | 3-bit Medium | ~6.2 GB | 8 GB | Balanced lightweight footprint for 8GB GPUs & laptops. |

| GPT-neo-Custom-14B-Q3_K_S.gguf | 3-bit Small | ~5.6 GB | 6–8 GB | Ultra-lightweight footprint for mobile and edge devices. |

| GPT-neo-Custom-14B-Q2_K.gguf | 2-bit | ~4.8 GB | 6 GB | Extreme compression for low-resource environments. |

---

🚀 Quickstart Deployment Guide

🦙 1. Ollama (Instant Run)

Run directly from Hugging Face:

ollama run hf.co/Szymon-Wachelko/GPT-neo-Custom-14B-GGUF

Or configure locally with a Modelfile:

FROM ./GPT-neo-Custom-14B-Q4_K_M.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER stop "<|im_end|>"
SYSTEM "Jesteś polskim asystentem AI GPT-neo-Custom 14B-DUS stworzonym przez Szymon-Wachelko. Zawsze odpowiadaj po polsku."

ollama create gpt-neo-custom-14b -f Modelfile
ollama run gpt-neo-custom-14b

---

⚡ 2. llama.cpp

Interactive Terminal Chat

llama-cli -m GPT-neo-Custom-14B-Q4_K_M.gguf -ngl 99 -cnv --color -p "Jesteś polskim asystentem AI GPT-neo-Custom 14B-DUS stworzonym przez Szymon-Wachelko. Zawsze odpowiadaj po polsku."

Standalone ChatGPT-like Web UI

llama-server -m GPT-neo-Custom-14B-Q4_K_M.gguf -ngl 99 --port 8080 -c 4096

> Open your browser at http://localhost:8080 for a full offline web interface.

---

🖥️ 3. LM Studio

1. Install LM Studio.
2. Download any .gguf file from the Files and versions tab above.
3. Drag and drop the .gguf file into LM Studio.
4. Set GPU Acceleration to MAX in the right panel.
5. Click New Chat and start conversation!

---

🐍 4. Python Integration

from llama_cpp import Llama

# Load model with full GPU offloading
llm = Llama(
    model_path="GPT-neo-Custom-14B-Q4_K_M.gguf",
    n_gpu_layers=-1, # -1 offloads all layers to GPU
    n_ctx=4096
)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "Jesteś asystentem AI GPT-neo-Custom 14B-DUS stworzonym przez Szymon-Wachelko. Zawsze odpowiadaj po polsku."},
        {"role": "user", "content": "Cześć! W czym możesz mi dzisiaj pomóc?"}
    ]
)

print(response["choices"][0]["message"]["content"])

---

💬 Prompt Template (ChatML)

The model is trained on the standard ChatML format:

<|im_start|>system
Jesteś asystentem AI GPT-neo-Custom 14B-DUS stworzonym przez Szymon-Wachelko. Zawsze odpowiadasz w języku polskim.<|im_end|>
<|im_start|>user
Napisz skrypt w Pythonie do pobierania kursów walut z API NBP.<|im_end|>
<|im_start|>assistant

---

💡 Community, Bug Reports & Feature Requests

The GPT-neo-Custom project is developed with an open-source mindset, and community feedback is vital for future iterations!

🛠️ Jak dodać pomysł lub zgłosić problem / How to contribute feedback:

1. Przejdź do zakładki Community / Discussions na samej górze tej strony.
2. Kliknij New Discussion.
3. Napisz nam:

* Błędy: Jakie pytania lub komendy sprawiają problem?

* Nowe funkcje: Co dodać w kolejnych wersjach? Jakie dziedziny wiedzy rozbudować?

* Wydajność: Jak model działa na Twojej karcie graficznej lub procesorze?

---

📜 Credits & License

• Base Model Architecture: Szymon-Wachelko/GPT-neo-Custom-7B-GGUF (Scaled from 28 &rarr; 56 layers via Depth Up-Scaling)
• Scaling Method: DUS (Depth Up-Scaling)
• Fine-Tuning & Quantization: Szymon-Wachelko
• License: Apache 2.0 (Free for personal, academic, and commercial use)