| | --- |
| | language: |
| | - tr |
| | - en |
| | - de |
| | - es |
| | - fr |
| | - ru |
| | - zh |
| | - ja |
| | - ko |
| | license: apache-2.0 |
| | tags: |
| | - turkish |
| | - türkiye |
| | - reasoning |
| | - vision-language |
| | - vlm |
| | - multimodal |
| | - lamapi |
| | - next2-air |
| | - qwen3.5 |
| | - text-generation |
| | - image-text-to-text |
| | - open-source |
| | - 2b |
| | - edge-ai |
| | - large-language-model |
| | - llm |
| | - thinking-mode |
| | - fast-inference |
| | 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: #0ea5e9; font-weight: 800; font-size: 2.8em; margin-bottom: 5px; letter-spacing: -1px;">Next2-Air (2B)</h1> |
| | <h3 style="color: #64748b; font-weight: 400; margin-top: 0; font-size: 1.2em;"><i>Türkiye’s Fastest Lightweight Multimodal & Reasoning AI</i></h3> |
| |
|
| | <p style="margin-top: 15px;"> |
| | <a href="https://opensource.org/licenses/Apache-2.0"><img src="https://img.shields.io/badge/License-Apache%202.0-blue.svg?style=for-the-badge" alt="License: Apache 2.0"></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-air"><img src="https://img.shields.io/badge/🤗_HuggingFace-Lamapi/Next2--Air-0ea5e9.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-Air** is a highly optimized, lightning-fast **2-Billion parameter Vision-Language Model (VLM)** built on the **Qwen 3.5-2B** architecture. Engineered by Lamapi in **Türkiye**, the "Air" moniker represents its core philosophy: **lightweight, incredibly fast, yet surprisingly capable.** |
| |
|
| | While large models dominate cloud servers, Next2-Air is designed to bring top-tier reasoning and multimodal understanding directly to your local machines, edge devices, and everyday applications. By utilizing specialized instruction-tuning and logical reasoning datasets, we have created a 2B model that thinks deeply, processes images flawlessly, and speaks native Turkish and English. |
| |
|
| | --- |
| |
|
| | ## ⚡ Highlights |
| |
|
| | <div style="background: #232323; border-left: 5px solid #0ea5e9; padding: 20px; width:fit-content; border-radius: 16px; font-family: sans-serif;"> |
| | <ul style="margin: 0; padding-left: 20px; line-height: 1.6; color: #808080;"> |
| | <li>🇹🇷 <strong>Perfected in Türkiye:</strong> Fine-tuned with cultural nuance, ensuring natural, fluent, and highly accurate Turkish responses.</li> |
| | <li>💨 <strong>"Air" Speed & Efficiency:</strong> Only 2 Billion parameters. Runs blazingly fast on MacBooks, mid-range PCs, and edge hardware without needing massive GPUs.</li> |
| | <li>🧠 <strong>Native Thinking Mode:</strong> Despite its small size, it leverages Chain-of-Thought (<code><think></code>) to logically deduce answers before speaking.</li> |
| | <li>👁️ <strong>Full Vision-Language Support:</strong> Analyzes images, reads documents (OCR), and understands visual context just like heavier models.</li> |
| | <li>📚 <strong>Massive Context:</strong> Supports a staggering <strong>262,144 tokens</strong> natively—perfect for summarizing long PDFs or reading extensive codebases locally.</li> |
| | </ul> |
| | </div> |
| | |
| | --- |
| |
|
| | ## 📊 Benchmark Performance |
| |
|
| | Next2-Air (2B) redefines what is possible in the ultra-lightweight category. Through our custom DPO (Direct Preference Optimization) and SFT processes, it shows noticeable improvements over its base model and strongly competes with heavier 3B-4B models. |
| |
|
| | ### 📝 Text, Reasoning & Instruction Following |
| |
|
| | <div style="overflow-x: auto; box-shadow: 0 4px 6px rgba(0,0,0,0.05); width:fit-content; border-radius: 16px;"> |
| | <table style="width: 100%; border-collapse: collapse; text-align: center; font-family: sans-serif; background: #232323; min-width: 800px;"> |
| | <thead> |
| | <tr style="background-color: #232323; color: white;"> |
| | <th style="padding: 14px; text-align: left; padding-left: 20px; border-radius: 16px 0 0 0;">Benchmark</th> |
| | <th style="padding: 14px; font-size: 1.1em;">Next2-Air (2B)</th> |
| | <th style="padding: 14px;">Qwen 3.5 (2B)</th> |
| | <th style="padding: 14px;">Gemma-2 (2B)</th> |
| | <th style="padding: 14px; border-radius: 0 16px 0 0;">Llama-3.2 (3B)</th> |
| | </tr> |
| | </thead> |
| | <tbody style="color: #808080;"> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: #232323; font-weight: 600;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;">MMLU-Pro (Thinking)</td> |
| | <td style="padding: 12px; color: #0ea5e9;">68.2%</td> |
| | <td style="padding: 12px;">66.5%</td> |
| | <td style="padding: 12px;">54.1%</td> |
| | <td style="padding: 12px;">68.4%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;;">MMLU-Redux</td> |
| | <td style="padding: 12px; font-weight: bold; color: #0ea5e9;">82.1%</td> |
| | <td style="padding: 12px;">79.6%</td> |
| | <td style="padding: 12px;">75.3%</td> |
| | <td style="padding: 12px;">79.5%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: #232323; font-weight: 600;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;">IFEval (Instruction)</td> |
| | <td style="padding: 12px; color: #0ea5e9;">82.5%</td> |
| | <td style="padding: 12px;">78.6%</td> |
| | <td style="padding: 12px;">75.8%</td> |
| | <td style="padding: 12px;">77.4%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;">TAU2-Bench (Agent)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #0ea5e9;">52.4%</td> |
| | <td style="padding: 12px;">48.8%</td> |
| | <td style="padding: 12px;">--</td> |
| | <td style="padding: 12px;">--</td> |
| | </tr> |
| | </tbody> |
| | </table> |
| | </div> |
| | |
| | ### 👁️ Multimodal & Vision Edge |
| |
|
| | Next2-Air features a highly capable visual encoder, allowing it to process spatial intelligence, OCR, and document understanding tasks efficiently. |
| |
|
| | <div style="overflow-x: auto; box-shadow: 0 4px 6px rgba(0,0,0,0.05); border-radius: 8px; margin-top: 15px;width:fit-content; "> |
| | <table style="width: 100%; border-collapse: collapse; text-align: center; font-family: sans-serif; background: #232323; min-width: 800px;"> |
| | <thead> |
| | <tr style="background-color: #232323; color: white;"> |
| | <th style="padding: 14px; text-align: left; padding-left: 20px; border-radius: 16px 0 0 0;">Benchmark</th> |
| | <th style="padding: 14px; font-size: 1.1em;">Next2-Air (2B)</th> |
| | <th style="padding: 14px; border-radius: 0 16px 0 0;">Base Qwen3.5-2B</th> |
| | </tr> |
| | </thead> |
| | <tbody style="color: #808080;"> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;;">MMMU (General VQA)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #0ea5e9;">66.5%</td> |
| | <td style="padding: 12px;">64.2%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: #232323;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;">MathVision</td> |
| | <td style="padding: 12px; font-weight: bold; color: #0ea5e9;">78.1%</td> |
| | <td style="padding: 12px;">76.7%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;">OCRBench</td> |
| | <td style="padding: 12px; font-weight: bold; color: #0ea5e9;">86.0%</td> |
| | <td style="padding: 12px;">84.5%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: #232323;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #0284c7;">VideoMME (w/ sub)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #0ea5e9;">77.8%</td> |
| | <td style="padding: 12px;">75.6%</td> |
| | </tr> |
| | </tbody> |
| | </table> |
| | </div> |
| | |
| | <p style="font-size: 0.85em; color: #888; margin-top: 10px;"><em>* Enhanced scores in reasoning and OCR are a direct result of Lamapi's specialized bilingual finetuning pipeline focusing on edge-case logic and structural formatting.</em></p> |
| |
|
| | --- |
| |
|
| | ## 🚀 Quickstart & Usage |
| |
|
| | **Next2-Air** is fully compatible with the Hugging Face `transformers` ecosystem and fast inference engines like `vLLM` and `SGLang`. Because it's a VLM, you can directly pass images into your prompts. |
| |
|
| | ### Python (Transformers) |
| |
|
| | Make sure you have `transformers`, `torch`, `torchvision`, and `pillow` installed. |
| |
|
| | ```python |
| | from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor |
| | from PIL import Image |
| | import torch |
| | |
| | model_id = "thelamapi/next2-air" |
| | |
| | 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 Air, 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 |
| |
|
| | | Attribute | Details | |
| | | :--- | :--- | |
| | | **Base Architecture** | Qwen 3.5 (Causal Language Model + Vision Encoder) | |
| | | **Parameters** | 2 Billion (Ultra-Lightweight) | |
| | | **Context Length** | 262,144 tokens natively | |
| | | **Hardware** | Optimized for Edge devices, MacBooks (MLX), Consumer GPUs, and low-VRAM environments. | |
| | | **Capabilities** | Text Generation, Image Understanding, OCR, Logic & Reasoning (CoT), Bilingual (TR/EN) | |
| |
|
| | --- |
| |
|
| | ## 🎯 Ideal Use Cases |
| |
|
| | **Next2-Air** is the undisputed champion of local, fast inference tasks. It is perfect for: |
| | * 🔋 **Mobile & Edge AI:** Deploying smart assistants natively on smartphones or Raspberry Pi without relying on cloud APIs. |
| | * ⚡ **Real-Time OCR & Parsing:** Quickly scanning receipts, invoices, or UI screenshots to extract data in milliseconds. |
| | * 💬 **Fast Conversational Bots:** Providing instant, low-latency Turkish and English responses for customer service pipelines. |
| | * 🎮 **Gaming & NPC Logic:** Acting as a fast reasoning engine for dynamic in-game characters. |
| |
|
| | --- |
| |
|
| | ## 📄 License & Open Source |
| |
|
| | Next2-Air is released under the **Apache 2.0 License**. We strongly believe in empowering developers, students, and enterprises with accessible, high-speed, reasoning-capable AI. |
| |
|
| | --- |
| |
|
| | ## 📞 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: 40px; padding: 25px; border-top: 1px solid #e0f2fe; background: #232323; border-radius: 8px;width:fit-content; "> |
| | <p style="color: #808080; font-size: 15px; margin: 0;"> |
| | <strong>Next2-Air</strong> — Hafif, Hızlı, Akıllı. Uç cihazlardan buluta, Türkiye'nin yeni nesil çevik yapay zekası. 🌬️ |
| | </p> |
| | </div> |
