| | --- |
| | language: |
| | - tr |
| | - en |
| | - de |
| | - es |
| | - fr |
| | - ru |
| | - zh |
| | - ja |
| | - ko |
| | license: mit |
| | tags: |
| | - turkish |
| | - türkiye |
| | - reasoning |
| | - ai |
| | - lamapi |
| | - next2 |
| | - next2-0.8b |
| | - qwen3.5 |
| | - text-generation |
| | - open-source |
| | - 0.8b |
| | - edge-ai |
| | - large-language-model |
| | - llm |
| | - transformer |
| | - artificial-intelligence |
| | - nlp |
| | - instruction-tuned |
| | - chat |
| | - thinking-mode |
| | - efficient |
| | - sft |
| | pipeline_tag: image-text-to-text |
| | datasets: |
| | - mlabonne/FineTome-100k |
| | - CognitiveKernel/CognitiveKernel-Pro-SFT |
| | - OpenSPG/KAG-Thinker-training-dataset |
| | - Gryphe/ChatGPT-4o-Writing-Prompts |
| | library_name: transformers |
| | --- |
| | |
| | <div align="center" style="font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;"> |
| | |
| | |
| |
|
| |  |
| |
|
| | <h1 style="color: #4A90E2; font-weight: 800; font-size: 2.5em; margin-bottom: 5px;">🧠 Next2 0.8B</h1> |
| | <h3 style="color: #888; font-weight: 400; margin-top: 0;"><i>Most Efficient & Compact Reasoning AI Model</i></h3> |
| |
|
| | <p> |
| | <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-blue.svg?style=for-the-badge" alt="License: MIT"></a> |
| | <a href="#"><img src="https://img.shields.io/badge/Language-TR%20%7C%20EN-red.svg?style=for-the-badge" alt="Language"></a> |
| | <a href="https://huggingface.co/Lamapi/next2-0.8b"><img src="https://img.shields.io/badge/🤗_HuggingFace-Lamapi/Next2--0.8B-orange.svg?style=for-the-badge" alt="HuggingFace"></a> |
| | <a href="https://discord.gg/XgH4EpyPD2"><img src="https://cdn-uploads.huggingface.co/production/uploads/67d46bc5fe6ad6f6511d6f44/NPUQziAExGvvY8exRUxw2.png" alt="Discord"></a> |
| | </p> |
| | |
| | </div> |
| |
|
| | --- |
| |
|
| | ## 📖 Overview |
| |
|
| | **Next2 0.8B** is a highly optimized, **800-million parameter** language model built on the cutting-edge **Qwen 3.5 architecture**. Carefully fine-tuned and developed in **Türkiye**, it is designed to deliver astonishing reasoning capabilities in a form factor small enough to run on local laptops, edge devices, and mobile environments. |
| |
|
| | Don't let the size fool you. Thanks to extensive **instruction tuning** and enhanced **Thinking Mode** datasets, Next2 0.8B punches significantly above its weight class. It introduces localized cultural nuances for Turkish users while maintaining top-tier English proficiency. It’s built to think, reason logically, and provide structured answers efficiently. |
| |
|
| | --- |
| |
|
| | ## ⚡ Highlights |
| |
|
| | <div style="background: rgba(74, 144, 226, 0.1); border-left: 4px solid #4A90E2; padding: 15px; border-radius: 4px;"> |
| | <ul> |
| | <li>🇹🇷 <strong>Developed & Fine-Tuned in Türkiye:</strong> Specially optimized for rich Turkish syntax and logical flows.</li> |
| | <li>🧠 <strong>Native Thinking Mode:</strong> Capable of chain-of-thought (CoT) reasoning for complex problem-solving.</li> |
| | <li>📱 <strong>Edge & Mobile Ready:</strong> At just 0.8B parameters, it runs blazingly fast on CPUs, low-end GPUs, and edge hardware.</li> |
| | <li>⚡ <strong>Enhanced Over Base:</strong> Noticeably improved mathematical reasoning and instruction following compared to standard 1B models.</li> |
| | </ul> |
| | </div> |
| | |
| | --- |
| |
|
| | ## 📊 Benchmark Performance |
| |
|
| | We tested **Next2 0.8B** against its base model and other models in the sub-2B category. Through careful dataset curation and SFT (Supervised Fine-Tuning) in Türkiye, it shows a tangible improvement in logical reasoning and contextual understanding. |
| |
|
| | <div style="overflow-x: auto;"> |
| | <table style="width: 100%; border-collapse: collapse; text-align: center; font-family: sans-serif;"> |
| | <thead> |
| | <tr style="background-color: #4A90E2; color: white;"> |
| | <th style="padding: 12px; border-radius: 8px 0 0 0;">Model</th> |
| | <th style="padding: 12px;">MMLU (5-shot)</th> |
| | <th style="padding: 12px;">IFEval</th> |
| | <th style="padding: 12px;">GSM8K (Math)</th> |
| | <th style="padding: 12px; border-radius: 0 8px 0 0;">Context Limit</th> |
| | </tr> |
| | </thead> |
| | <tbody> |
| | <tr style="background-color: rgba(74, 144, 226, 0.05); font-weight: bold; border-bottom: 1px solid #ddd;"> |
| | <td style="padding: 10px; color: #4A90E2;">🚀 Next2 0.8B (Thinking)</td> |
| | <td style="padding: 10px;">52.1%</td> |
| | <td style="padding: 10px;">55.8%</td> |
| | <td style="padding: 10px;">67.4%</td> |
| | <td style="padding: 10px;">32K+</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #ddd;"> |
| | <td style="padding: 10px;">Base Qwen3.5-0.8B</td> |
| | <td style="padding: 10px;">48.5%</td> |
| | <td style="padding: 10px;">52.1%</td> |
| | <td style="padding: 10px;">62.2%</td> |
| | <td style="padding: 10px;">262K</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #ddd;"> |
| | <td style="padding: 10px;">Llama-3.2-1B</td> |
| | <td style="padding: 10px;">49.3%</td> |
| | <td style="padding: 10px;">50.2%</td> |
| | <td style="padding: 10px;">60.5%</td> |
| | <td style="padding: 10px;">128K</td> |
| | </tr> |
| | </tbody> |
| | </table> |
| | </div> |
| | <p style="font-size: 0.85em; color: #666; margin-top: 10px;"><em>* Scores represent generalized task performance. Next2 0.8B shows a distinct advantage in reasoning (GSM8K) and instruction following (IFEval) due to our proprietary fine-tuning pipelines.</em></p> |
| | |
| | --- |
| |
|
| | ## 🚀 Quickstart & Usage |
| |
|
| | You can easily run **Next2 0.8B** on almost any machine with Python installed. Because of its size, `device_map="auto"` will comfortably map it to memory without breaking a sweat. |
| |
|
| | ```python |
| | from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor |
| | from PIL import Image |
| | import torch |
| | |
| | model_id = "thelamapi/next2-0.8b" |
| | |
| | model = AutoModelForCausalLM.from_pretrained(model_id) |
| | processor = AutoProcessor.from_pretrained(model_id) # For vision. |
| | tokenizer = AutoTokenizer.from_pretrained(model_id) |
| | |
| | |
| | # Create a message in chat format |
| | messages = [ |
| | {"role": "system","content": [{"type": "text", "text": "You are Next2, a smart and concise AI assistant trained by Lamapi. Always respond in the user's language. Proudly made in Turkey."}]}, |
| | |
| | { |
| | "role": "user","content": [ |
| | {"type": "text", "text": "Write a highly optimized Rust function to calculate the Fibonacci sequence using memoization"} |
| | ] |
| | } |
| | ] |
| | |
| | # Prepare input with Tokenizer |
| | prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False) |
| | inputs = processor(text=prompt, return_tensors="pt") |
| | |
| | # Remove 'mm_token_type_ids' if it's not needed for text-only generation |
| | if "mm_token_type_ids" in inputs: |
| | del inputs["mm_token_type_ids"] |
| | |
| | |
| | # Output from the model |
| | output = model.generate(**inputs, do_sample=True, temperature=0.7, max_new_tokens=128) |
| | print(tokenizer.decode(output[0], skip_special_tokens=True)) |
| | ``` |
| |
|
| | --- |
| |
|
| | ## 🧩 Model Specifications |
| |
|
| | | Feature | Details | |
| | | :--- | :--- | |
| | | **Base Architecture** | Qwen 3.5 (Transformer with Gated Delta Networks) | |
| | | **Parameter Count** | 0.8 Billion (800M) | |
| | | **Primary Focus** | Edge Inference, Reasoning (CoT), Turkish/English Bilingual | |
| | | **Optimizations** | Multi-Token Prediction (MTP) Support, Flash Attention ready | |
| | | **Hardware Reqs** | Ultra-lightweight (Can run on 2GB RAM / Edge GPUs) | |
| | | **Format** | FP16 natively, Quantization (GGUF/AWQ) recommended for mobile | |
| |
|
| | --- |
| |
|
| | ## 🎯 Ideal Use Cases |
| |
|
| | Since it is compact yet surprisingly capable, Next2 0.8B is perfect for: |
| | * 🔋 **On-Device AI:** Running locally on smartphones, Raspberry Pi, or older laptops without internet. |
| | * 🤖 **NPC & Gaming AI:** Fast, low-latency dialogue generation for video games. |
| | * 📝 **Text Summarization & Extraction:** Processing documents locally to maintain high data privacy. |
| | * 🇹🇷 **Turkish NLP Tasks:** Fast classification, sentiment analysis, and daily conversational AI in Turkish. |
| |
|
| | --- |
| |
|
| | ## 📄 License & Open Source |
| |
|
| | Licensed under the **MIT License**. We believe in democratizing AI, making smart, reasoning-capable models accessible to everyone. Feel free to use it in commercial apps, academic research, or personal projects! |
| |
|
| | --- |
| |
|
| | ## 📞 Contact & Community |
| |
|
| | * 📧 **Email:** [lamapicontact@gmail.com](mailto:lamapicontact@gmail.com) |
| | * 🤗 **HuggingFace:** [Lamapi](https://huggingface.co/Lamapi) |
| | * 💬 **Discord:** [Join the Lamapi Community](https://discord.gg/XgH4EpyPD2) |
| |
|
| | --- |
| |
|
| | <div align="center" style="margin-top: 30px; padding: 20px; border-top: 1px solid #eaeaea;"> |
| | <p style="color: #666; font-size: 14px;"> |
| | <strong>Next2 0.8B</strong> — Küçük boyutlu, büyük akıllı. Türkiye'den dünyaya, sınır tanımayan yeni nesil yerel AI. 🌍 |
| | </p> |
| | </div> |
