# 🚀 Next 4B (s330) ### *Türkiye’s 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) --- ## 📖 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 **Türkiye’s 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.

Model	MMLU (5-shot) %	MMLU-Pro %	GSM8K %	MATH %
Next 4B preview Version s325	84.6	66.9	82.7	70.5
Next 1B Version t327	87.3	69.2	90.5	70.1
Qwen 3 0.6B	52.81	37.6	60.7	20.5
Llama 3.2 1B	49.3	44.4	11.9	30.6

--- # Also, our Next Z1 model is leading to state-of-the-art models in some of the Benchmarks.

Model	MMLU (5-shot) %	MMLU-Pro %	GSM8K %	MATH %
Next Z1 Version l294	97.3	94.2	97.7	93.2
Next Z1 Version l294 (no tool)	94.7	90.1	94.5	88.7
GPT 5	92.5	87.0	98.4	96.0
Claude Opus 4.1 (Thinking)	~92.0	87.8	84.7	95.4

--- ## 🚀 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)) ```

Who is in this image?

The image shows Mustafa Kemal Atatürk, the founder and first President of the Republic of Turkey.

### 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)) ```

Hello, how are you?

I'm fine, thank you. How are you?

--- ## 🎯 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** — Türkiye’s **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)