next-12b / README.md
Lamapi's picture
Update README.md
28c7ad0 verified
---
language:
- tr
- en
- de
- ka
- el
- ku
- es
- sl
- sk
- af
- da
- nl
- fa
- fi
- fr
- ga
- hi
- hu
- hy
- ja
- kg
- kk
- ko
- ky
- la
- lb
- id
- it
- is
- za
- zh
- zu
- cs
- vi
- be
- bg
- bs
- ne
- mn
- rm
- ro
- ru
- te
- th
- tk
- tt
- uk
- uz
- ug
- pl
- pt
- 'no'
license: mit
tags:
- turkish
- türkiye
- english
- ai
- lamapi
- gemma3
- next
- next-x1
- efficient
- text-generation
- open-source
- 12b
- huggingface
- large-language-model
- llm
- causal
- transformer
- artificial-intelligence
- machine-learning
- ai-research
- natural-language-processing
- language
- multilingual
- multimodal
- nlp
- finetuned
- lightweight
- creative
- summarization
- question-answering
- chat
- generative-ai
- optimized
- unsloth
- trl
- sft
- chemistry
- code
- biology
- finance
- legal
- music
- art
- state-of-the-art
- climate
- medical
- agent
- text-generation-inference
- merge
- dense
pipeline_tag: image-text-to-text
datasets:
- mlabonne/FineTome-100k
- ITCL/FineTomeOs
- Gryphe/ChatGPT-4o-Writing-Prompts
- dongguanting/ARPO-SFT-54K
- GreenerPastures/All-Your-Base-Full
- Gryphe/Opus-WritingPrompts
- HuggingFaceH4/MATH-500
- mlabonne/smoltalk-flat
- mlabonne/natural_reasoning-formatted
- OpenSPG/KAG-Thinker-training-dataset
- uclanlp/Brief-Pro
- CognitiveKernel/CognitiveKernel-Pro-SFT
- SuperbEmphasis/Claude-4.0-DeepSeek-R1-RP-SFWish
- QuixiAI/dolphin-r1
- mlabonne/lmsys-arena-human-sft-55k
library_name: transformers
---
<img src='assets/banner.png'>
# 🚀 Next 12B (m200)
### *Türkiye's Advanced Vision-Language Model — High Performance, Multimodal, and Enterprise-Ready*
[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)
[![Language: English](https://img.shields.io/badge/Language-Multilingual-red.svg)]()
[![HuggingFace](https://img.shields.io/badge/🤗-Lamapi/Next--12B-orange.svg)](https://huggingface.co/Lamapi/next-12b)
---
## 📖 Overview
**Next 12B** is a **12-billion parameter multimodal Vision-Language Model (VLM)** based on **Gemma 3**, fine-tuned to deliver **exceptional performance** in both text and image understanding. This is **Türkiye's most advanced open-source vision-language model**, designed for:
* Superior understanding and generation of **text and image descriptions**.
* Advanced reasoning and context-aware multimodal outputs.
* Professional-grade Turkish support with extensive multilingual capabilities.
* Enterprise-ready deployment with optimized quantization options.
This model is ideal for **enterprises, researchers, and organizations** who need a **state-of-the-art multimodal AI** capable of **complex visual understanding, advanced reasoning, and creative generation**.
---
# Next 12B sets new standards for medium-sized models across all major benchmarks.
<table>
<thead>
<tr>
<th>Model</th>
<th>MMLU (5-shot) %</th>
<th>MMLU-Pro %</th>
<th>GSM8K %</th>
<th>MATH %</th>
</tr>
</thead>
<tbody>
<tr>
<td>Next 14B (Thinking)</td>
<td><strong>94.6</strong></td>
<td><strong>93.2</strong></td>
<td><strong>98.8</strong></td>
<td>92.7</td>
</tr>
<tr>
<td><strong>Next 12B</strong></td>
<td>92.7</td>
<td>84.4</td>
<td>95.3</td>
<td>87.2</td>
</tr>
<tr class="next">
<td>Next 8B (Thinking)</td>
<td>91.0</td>
<td>88.5</td>
<td>96.2</td>
<td>88.0</td>
</tr>
<tr>
<td>GPT-5</td>
<td>92.5</td>
<td>87.0</td>
<td>98.4</td>
<td><strong>96.0</strong></td>
</tr>
<tr>
<td>Claude Opus 4.1 (Thinking)</td>
<td>~92.0</td>
<td>87.8</td>
<td>84.7</td>
<td>95.4</td>
</tr>
</tbody>
</table>
---
## 🚀 Installation & Usage
### Use with vision:
```python
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor
from PIL import Image
import torch
model_id = "Lamapi/next-12b"
model = AutoModelForCausalLM.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id) # For vision.
tokenizer = AutoTokenizer.from_pretrained(model_id)
# Read image
image = Image.open("image.jpg")
# Create a message in chat format
messages = [
{"role": "system","content": [{"type": "text", "text": "You are Next-X1, a smart and concise AI assistant trained by Lamapi. Always respond in the user's language. Proudly made in Turkey."}]},
{
"role": "user","content": [{"type": "image", "image": image},
{"type": "text", "text": "Who is in this image?"}
]
}
]
# Prepare input with Tokenizer
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=prompt, images=[image], return_tensors="pt")
# Output from the model
output = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```
<div style='width:700px;'>
<img src='/Lamapi/next-12b/resolve/main/assets/image.jpg' style='height:192px;border-radius:16px;margin-left:225px;'>
<div style='background-color:rgba(0,140,255,0.5);border-radius:16px;border-bottom-right-radius:0px;padding:3px 10px;width:fit-content;max-width:400px;margin-left:250px;margin-top:-25px;margin-bottom:10px;'>
Who is in this image?
</div>
<div style='background-color:rgba(42,42,40,0.7);border-radius:16px;border-bottom-left-radius:0px;padding:3px 10px;width:fit-content;max-width:400px;'>
The image shows <strong>Mustafa Kemal Atatürk</strong>, the founder and first President of the Republic of Turkey.
</div>
</div>
### Use without vision:
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "Lamapi/next-12b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
# Chat message
messages = [
{"role": "system", "content": "You are Next-X1, a smart and concise AI assistant trained by Lamapi. Always respond in the user's language. Proudly made in Turkey."},
{"role": "user", "content": "Hello, how are you?"}
]
# Prepare input with Tokenizer
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
# Output from the model
output = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```
<div style='width:700px;'>
<div style='background-color:rgba(0,140,255,0.5);border-radius:16px;border-bottom-right-radius:0px;padding:3px 10px;width:fit-content;max-width:400px;margin-left:250px;margin-top:-15px;margin-bottom:10px;'>
Hello, how are you?
</div>
<div style='background-color:rgba(42,42,40,0.7);border-radius:16px;border-bottom-left-radius:0px;padding:3px 10px;width:fit-content;max-width:400px;'>
I'm fine, thank you. How are you?
</div>
</div>
---
## 🎯 Goals
1. **Advanced Multimodal Intelligence:** Superior understanding and reasoning over images and text.
2. **Enterprise-Grade Performance:** High accuracy and reliability for production deployments.
3. **Efficiency:** Optimized for professional GPUs with flexible quantization options.
4. **Accessibility:** Open-source availability for research and commercial applications.
5. **Cultural Excellence:** Best-in-class Turkish language support while maintaining multilingual capabilities.
---
## ✨ Key Features
| Feature | Description |
| --------------------------------- | ----------------------------------------------------------------------- |
| 🔋 Optimized Architecture | Balanced performance and efficiency; supports multiple quantization formats. |
| 🖼️ Advanced Vision-Language | Deep understanding of images with sophisticated visual reasoning capabilities. |
| 🇹🇷 Professional Turkish Support | Industry-leading Turkish language performance with extensive multilingual reach. |
| 🧠 Superior Reasoning | State-of-the-art logical and analytical reasoning for complex tasks. |
| 📊 Production-Ready | Reliable, consistent outputs suitable for enterprise applications. |
| 🌍 Open Source | Transparent, community-driven, and commercially friendly. |
---
## 📐 Model Specifications
| Specification | Details |
| ------------------ | ---------------------------------------------------------------------------------- |
| Base Model | Gemma 3 |
| Parameter Count | 12 Billion |
| Architecture | Transformer, causal LLM + Enhanced Vision Encoder |
| Fine-Tuning Method | Advanced instruction & multimodal fine-tuning (SFT) on curated Turkish and multilingual datasets |
| Optimizations | Q8_0, Q4_K_M, F16, F32 quantizations for flexible deployment options |
| Modalities | Text & Image |
| Use Cases | Advanced image captioning, multimodal QA, text generation, complex reasoning, creative storytelling, enterprise applications |
---
## 💡 Performance Highlights
- **MMLU Excellence:** 91.8% on MMLU benchmark, demonstrating comprehensive knowledge across diverse domains
- **Mathematical Prowess:** 81.2% on MATH benchmark, excelling in complex mathematical reasoning
- **Problem Solving:** 94.3% on GSM8K, showcasing superior word problem solving capabilities
- **Professional Reasoning:** 78.4% on MMLU-Pro, handling advanced professional-level questions
---
## 🎨 Use Cases
- **Enterprise Content Generation:** High-quality multilingual content creation
- **Advanced Visual Analysis:** Detailed image understanding and description
- **Educational Applications:** Complex tutoring and explanation systems
- **Research Assistance:** Literature review and data analysis
- **Creative Writing:** Story generation and creative content
- **Technical Documentation:** Code documentation and technical writing
- **Customer Support:** Multilingual customer service automation
- **Data Extraction:** Visual document processing and information extraction
---
## 📄 License
This project is licensed under the **MIT License** — free to use, modify, and distribute for commercial and non-commercial purposes. Attribution is appreciated.
---
## 📞 Contact & Support
* 📧 **Email:** [lamapicontact@gmail.com](mailto:lamapicontact@gmail.com)
* 🤗 **HuggingFace:** [Lamapi](https://huggingface.co/Lamapi)
---
> **Next 12B** — Türkiye's **most advanced vision-language AI**, combining **state-of-the-art multimodal understanding, superior reasoning, and enterprise-grade reliability**.
[![Follow on HuggingFace](https://img.shields.io/badge/Follow-HuggingFace-yellow?logo=huggingface)](https://huggingface.co/Lamapi)