248 lines
7.5 KiB
Markdown
248 lines
7.5 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
- tr
|
|
- de
|
|
- fr
|
|
- es
|
|
- it
|
|
- pt
|
|
- ru
|
|
- zh
|
|
- ja
|
|
- ko
|
|
- hi
|
|
- ar
|
|
- nl
|
|
- pl
|
|
- uk
|
|
- vi
|
|
- th
|
|
- id
|
|
- cs
|
|
license: mit
|
|
tags:
|
|
- global-ai
|
|
- multilingual
|
|
- vision-language-model
|
|
- multimodal
|
|
- lamapi
|
|
- next-2-fast
|
|
- next-series
|
|
- 4b
|
|
- efficient
|
|
- gemma-3
|
|
- transformer
|
|
- text-generation
|
|
- reasoning
|
|
- artificial-intelligence
|
|
- nlp
|
|
pipeline_tag: image-text-to-text
|
|
datasets:
|
|
- mlabonne/FineTome-100k
|
|
- ITCL/FineTomeOs
|
|
- Gryphe/ChatGPT-4o-Writing-Prompts
|
|
- dongguanting/ARPO-SFT-54K
|
|
- OpenSPG/KAG-Thinker-training-dataset
|
|
- uclanlp/Brief-Pro
|
|
- CognitiveKernel/CognitiveKernel-Pro-SFT
|
|
- QuixiAI/dolphin-r1
|
|
library_name: transformers
|
|
---
|
|
|
|
|
|
|
|
[](https://discord.gg/XgH4EpyPD2)
|
|
|
|
# ⚡ Next 2 Fast (4B)
|
|
|
|
### *Global Speed, Multimodal Intelligence — Engineered by Lamapi*
|
|
|
|
[](https://opensource.org/licenses/MIT)
|
|
[]()
|
|
[](https://huggingface.co/Lamapi/next-2-fast)
|
|
|
|
---
|
|
|
|
## 🌍 Overview
|
|
|
|
**Next 2 Fast** is a state-of-the-art **4-billion parameter Multimodal Vision-Language Model (VLM)** designed for high-performance reasoning across languages and modalities.
|
|
|
|
Developed by **Lamapi**, a leading AI research lab in Türkiye, this model represents a leap in efficiency, bridging the gap between massive commercial models and accessible, open-source intelligence. Built upon the **Gemma 3** architecture and refined with our proprietary SFT and DPO techniques, **Next 2 Fast** is not just a language model—it is a global reasoning engine that sees, understands, and communicates fluently in **English, Turkish, German, French, Spanish, and 25+ other languages.**
|
|
|
|
**Why Next 2 Fast?**
|
|
* ⚡ **Global Performance:** Tuned for complex reasoning in English and multilingual contexts, outperforming larger models.
|
|
* 👁️ **Vision & Text:** Seamlessly processes images and text to generate code, descriptions, and analysis.
|
|
* 🚀 **Unmatched Speed:** Optimized for low-latency inference, making it ~2x faster than previous generations.
|
|
* 🔋 **Efficient Deployment:** Runs smoothly on consumer hardware (8GB VRAM) using 4-bit/8-bit quantization.
|
|
|
|
---
|
|
|
|
# 🏆 Benchmark Performance
|
|
|
|
**Next 2 Fast** delivers flagship-level performance in a compact 4B size, proving that efficiency does not require sacrificing intelligence.
|
|
|
|
<table>
|
|
<thead>
|
|
<tr>
|
|
<th>Model</th>
|
|
<th>Params</th>
|
|
<th>MMLU (5-shot) %</th>
|
|
<th>MMLU-Pro %</th>
|
|
<th>GSM8K %</th>
|
|
<th>MATH %</th>
|
|
</tr>
|
|
</thead>
|
|
<tbody>
|
|
<tr class="next" style="background-color: #e6f3ff; font-weight: bold;">
|
|
<td data-label="Model">⚡ Next 2 Fast</td>
|
|
<td>4B</td>
|
|
<td data-label="MMLU (5-shot) %">85.1</td>
|
|
<td data-label="MMLU-Pro %">67.4</td>
|
|
<td data-label="GSM8K %">83.5</td>
|
|
<td data-label="MATH %"><strong>71.2</strong></td>
|
|
</tr>
|
|
<tr>
|
|
<td data-label="Model">Gemma 3 4B</td>
|
|
<td>4B</td>
|
|
<td data-label="MMLU (5-shot) %">82.0</td>
|
|
<td data-label="MMLU-Pro %">64.5</td>
|
|
<td data-label="GSM8K %">80.1</td>
|
|
<td data-label="MATH %">68.0</td>
|
|
</tr>
|
|
<tr>
|
|
<td data-label="Model">Llama 3.2 3B</td>
|
|
<td>3B</td>
|
|
<td data-label="MMLU (5-shot) %">63.4</td>
|
|
<td data-label="MMLU-Pro %">52.1</td>
|
|
<td data-label="GSM8K %">45.2</td>
|
|
<td data-label="MATH %">42.8</td>
|
|
</tr>
|
|
<tr>
|
|
<td data-label="Model">Phi-3.5 Mini</td>
|
|
<td>3.8B</td>
|
|
<td data-label="MMLU (5-shot) %">84.0</td>
|
|
<td data-label="MMLU-Pro %">66.0</td>
|
|
<td data-label="GSM8K %">82.0</td>
|
|
<td data-label="MATH %">69.5</td>
|
|
</tr>
|
|
</tbody>
|
|
</table>
|
|
|
|
---
|
|
|
|
## 🚀 Quick Start
|
|
|
|
**Next 2 Fast** is fully compatible with the Hugging Face `transformers` library.
|
|
|
|
### 🖼️ Multimodal Inference (Vision + Text):
|
|
|
|
```python
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor
|
|
from PIL import Image
|
|
import torch
|
|
|
|
model_id = "thelamapi/next2-fast"
|
|
|
|
# Load Model & Processor
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_id,
|
|
torch_dtype=torch.bfloat16,
|
|
device_map="auto"
|
|
)
|
|
processor = AutoProcessor.from_pretrained(model_id)
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
|
|
# Load Image
|
|
image = Image.open("image.jpg")
|
|
|
|
# Create Multimodal Prompt
|
|
messages = [
|
|
{
|
|
"role": "system",
|
|
"content": [{"type": "text", "text": "You are Next-2, an AI assistant created by Lamapi. Provide concise and accurate analysis."}]
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": [
|
|
{"type": "image", "image": image},
|
|
{"type": "text", "text": "Analyze this image and explain in English."}
|
|
]
|
|
}
|
|
]
|
|
|
|
# Process & Generate
|
|
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|
inputs = processor(text=prompt, images=[image], return_tensors="pt").to(model.device)
|
|
|
|
output = model.generate(**inputs, max_new_tokens=128)
|
|
print(tokenizer.decode(output[0], skip_special_tokens=True))
|
|
```
|
|
|
|
### 💬 Text-Only Chat (Global Reasoning):
|
|
|
|
```python
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
import torch
|
|
|
|
model_id = "Lamapi/next-2-fast"
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_id,
|
|
torch_dtype=torch.bfloat16,
|
|
device_map="auto"
|
|
)
|
|
|
|
messages = [
|
|
{"role": "system", "content": "You are Next 2 Fast, an advanced AI assistant."},
|
|
{"role": "user", "content": "Explain the concept of entropy in thermodynamics simply."}
|
|
]
|
|
|
|
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
|
|
|
output = model.generate(**inputs, max_new_tokens=200)
|
|
print(tokenizer.decode(output[0], skip_special_tokens=True))
|
|
```
|
|
|
|
---
|
|
|
|
## 🌐 Key Features
|
|
|
|
| Feature | Description |
|
|
| :--- | :--- |
|
|
| **🌍 True Multilingualism** | Fluent in English, Turkish, German, French, Spanish, and more. No "translation-ese." |
|
|
| **🧠 Visual Intelligence** | Can read charts, identify objects, and reason about visual scenes effectively. |
|
|
| **⚡ High Efficiency** | Designed for speed. Ideal for edge devices, local deployment, and real-time apps. |
|
|
| **💻 Code & Math** | Strong capabilities in Python coding, debugging, and solving mathematical problems. |
|
|
| **🛡️ Global Alignment** | Fine-tuned with a diverse dataset to ensure safety and neutrality across cultures. |
|
|
|
|
---
|
|
|
|
## 🎯 Mission
|
|
|
|
At **Lamapi**, our mission is to build the **Next** generation of intelligence that is accessible to everyone, everywhere.
|
|
|
|
**Next 2 Fast** proves that world-class AI innovation isn't limited to Silicon Valley. By combining efficient architecture with high-quality global datasets, we provide a powerful tool for researchers, developers, and businesses worldwide.
|
|
|
|
---
|
|
|
|
## 📄 License
|
|
|
|
This model is open-sourced under the **MIT License**. It is free for academic and commercial use.
|
|
|
|
---
|
|
|
|
## 📞 Contact & Ecosystem
|
|
|
|
We are **Lamapi**.
|
|
|
|
* 📧 **Contact:** [Mail](mailto:lamapicontact@gmail.com)
|
|
* 🤗 **HuggingFace:** [Company Page](https://huggingface.co/thelamapi)
|
|
|
|
---
|
|
|
|
> **Next 2 Fast** — *Global Intelligence. Lightning Speed. Powered by Lamapi.*
|
|
|
|
[](https://huggingface.co/Lamapi) |