Files
next-4b/README.md
ModelHub XC e8c3c6eed7 初始化项目,由ModelHub XC社区提供模型
Model: thelamapi/next-4b
Source: Original Platform
2026-08-26 19:43:18 +08:00

382 lines
11 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
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
- 4b
- 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 4B (s330)
### *Türkiyes First Vision-Language Model — Efficient, Multimodal, and Reasoning-Focused*
[![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--4B-orange.svg)](https://huggingface.co/Lamapi/next-4b)
[![Discord](https://cdn.modrinth.com/data/cached_images/e84c69448cbf878a167f996d63e1a253437fcea2.png)](https://discord.gg/XgH4EpyPD2)
---
## 📖 Overview
**Next 4B** is a **4-billion parameter multimodal Vision-Language Model (VLM)** based on **Gemma 3**, fine-tuned to handle **both text and images** efficiently. It is **Türkiyes first open-source vision-language model**, designed for:
* Understanding and generating **text and image descriptions**.
* Efficient reasoning and context-aware multimodal outputs.
* Turkish support with multilingual capabilities.
* Low-resource deployment using **8-bit quantization** for consumer-grade GPUs.
This model is ideal for **researchers, developers, and organizations** who need a **high-performance multimodal AI** capable of **visual understanding, reasoning, and creative generation**.
---
# Our Next 1B and Next 4B models are leading to all of the tiny models in 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 class="next">
<td data-label="Model">Next 4B preview</td>
<td data-label="MMLU (5-shot) %">84.6</td>
<td data-label="MMLU-Pro %">66.9</td>
<td data-label="GSM8K %">82.7</td>
<td data-label="MATH %"><strong>70.5</strong></td>
</tr>
<tr class="next">
<td data-label="Model">Next 1B</td>
<td data-label="MMLU (5-shot) %"><strong>87.3</strong></td>
<td data-label="MMLU-Pro %"><strong>69.2</strong></td>
<td data-label="GSM8K %"><strong>90.5</strong></td>
<td data-label="MATH %">70.1</td>
</tr>
<tr>
<td data-label="Model">Qwen 3 0.6B</td>
<td data-label="MMLU (5-shot) %">52.81</td>
<td data-label="MMLU-Pro %">37.6</td>
<td data-label="GSM8K %">60.7</td>
<td data-label="MATH %">20.5</td>
</tr>
<tr>
<td data-label="Model">Llama 3.2 1B</td>
<td data-label="MMLU (5-shot) %">49.3</td>
<td data-label="MMLU-Pro %">44.4</td>
<td data-label="GSM8K %">11.9</td>
<td data-label="MATH %">30.6</td>
</tr>
</tbody>
</table>
---
# Also, our Next 14b model is leading to state-of-the-art models in some of the 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 class="next">
<td><strong>Next 14B (Thinking)</strong></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>Next 12B</td>
<td>92.7</td>
<td>84.4</td>
<td>95.3</td>
<td>87.2</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-4b"
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-4b/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-4b"
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. **Multimodal Intelligence:** Understand and reason over images and text.
2. **Efficiency:** Run on modest GPUs using 8-bit quantization.
3. **Accessibility:** Open-source availability for research and applications.
4. **Cultural Relevance:** Optimized for Turkish language and context while remaining multilingual.
---
## ✨ Key Features
| Feature | Description |
| --------------------------------- | ----------------------------------------------------------------------- |
| 🔋 Efficient Architecture | Optimized for low VRAM; supports 8-bit quantization for consumer GPUs. |
| 🖼️ Vision-Language Capable | Understands images, captions them, and performs visual reasoning tasks. |
| 🇹🇷 Multilingual & Turkish-Ready | Handles complex Turkish text with high accuracy. |
| 🧠 Advanced Reasoning | Supports logical and analytical reasoning for both text and images. |
| 📊 Consistent & Reliable Outputs | Reproducible responses across multiple runs. |
| 🌍 Open Source | Transparent, community-driven, and research-friendly. |
---
## 📐 Model Specifications
| Specification | Details |
| ------------------ | ---------------------------------------------------------------------------------- |
| Base Model | Gemma 3 |
| Parameter Count | 4 Billion |
| Architecture | Transformer, causal LLM + Vision Encoder |
| Fine-Tuning Method | Instruction & multimodal fine-tuning (SFT) on Turkish and multilingual datasets |
| Optimizations | Q8_0, F16, F32 quantizations for low VRAM and high VRAM usage |
| Modalities | Text & Image |
| Use Cases | Image captioning, multimodal QA, text generation, reasoning, creative storytelling |
---
## 📄 License
This project is licensed under the **MIT License** — free to use, modify, and distribute. Attribution is appreciated.
---
## 📞 Contact & Support
* 📧 **Email:** [lamapicontact@gmail.com](mailto:lamapicontact@gmail.com)
* 🤗 **HuggingFace:** [Lamapi](https://huggingface.co/Lamapi)
---
> **Next 4B** — Türkiyes **first vision-language AI**, combining **multimodal understanding, reasoning, and efficiency**.
[![Follow on HuggingFace](https://img.shields.io/badge/Follow-HuggingFace-yellow?logo=huggingface)](https://huggingface.co/Lamapi)