| | --- |
| | language: |
| | - tr |
| | - en |
| | - de |
| | - es |
| | - fr |
| | - ru |
| | - zh |
| | - ja |
| | - ko |
| | license: apache-2.0 |
| | tags: |
| | - turkish |
| | - türkiye |
| | - reasoning |
| | - vision-language |
| | - vlm |
| | - multimodal |
| | - lamapi |
| | - next2.5 |
| | - qwen3.5 |
| | - gemma-3 |
| | - text-generation |
| | - image-text-to-text |
| | - open-source |
| | - 4b |
| | - edge-ai |
| | - large-language-model |
| | - llm |
| | - thinking-mode |
| | 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: #6366F1; font-weight: 800; font-size: 2.8em; margin-bottom: 5px; letter-spacing: -1px;">🚀 Next 2.5 (4B)</h1> |
| | <h3 style="color: #64748b; font-weight: 400; margin-top: 0; font-size: 1.2em;"><i>Türkiye’s Advanced Native 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/next2.5-4b"><img src="https://img.shields.io/badge/🤗_HuggingFace-Lamapi/Next2.5--4B-indigo.svg?style=for-the-badge&color=6366F1" 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 |
| |
|
| | **Next 2.5** is a state-of-the-art **4-Billion parameter Vision-Language Model (VLM)**, built upon the powerful **Qwen 3.5-4B** foundation. Developed and heavily fine-tuned in **Türkiye** by Lamapi, Next 2.5 pushes the boundaries of what mid-sized models can achieve in 2026. |
| |
|
| | We have taken the already exceptional multimodal and reasoning capabilities of the base model and supercharged them through customized instruction tuning, culturally aware Turkish datasets, and enhanced visual-spatial reasoning tasks. Next 2.5 is designed to "think before it speaks", seamlessly analyzing complex images, videos, and intricate mathematical problems natively. |
| |
|
| | --- |
| |
|
| | ## ⚡ Highlights |
| |
|
| | <div style="background: linear-gradient(145deg, rgba(99, 102, 241, 0.05), rgba(99, 102, 241, 0.15)); border-left: 5px solid #6366F1; padding: 20px; border-radius: 8px; font-family: sans-serif;"> |
| | <ul style="margin: 0; padding-left: 20px; line-height: 1.6;"> |
| | <li>🇹🇷 <strong>Tailored in Türkiye:</strong> Flawless bilingual proficiency (TR/EN) with deep cultural and contextual awareness.</li> |
| | <li>🧠 <strong>Native Thinking Mode:</strong> By default, it uses <code><think>...</think></code> blocks to reason through complex logic, math, and coding tasks before answering.</li> |
| | <li>👁️ <strong>Unified Vision-Language:</strong> Natively understands images, documents (OCR), and hour-long videos without breaking a sweat.</li> |
| | <li>📈 <strong>Class-Leading Performance:</strong> Outperforms heavyweights in its parameter class (Gemma-3-4B, Phi-4-Mini) and even rivals closed-source edge models like GPT-5-Nano.</li> |
| | <li>📚 <strong>Massive Context Limit:</strong> Supports up to <strong>262,144 tokens</strong> natively, perfect for massive codebases or multi-document analysis.</li> |
| | </ul> |
| | </div> |
| | |
| | --- |
| |
|
| | ## 📊 Comprehensive Benchmarks |
| |
|
| | Through rigorous SFT and DPO phases, **Next 2.5 (4B)** sets a new standard for the ~4B parameter weight class. It consistently outperforms modern edge models and punches far above its weight, rivaling 8B-11B models in vision and reasoning. |
| |
|
| | ### 📝 Text, Knowledge & Reasoning (Sub-5B Class) |
| |
|
| | <div style="overflow-x: auto; box-shadow: 0 4px 6px rgba(0,0,0,0.05); border-radius: 8px;"> |
| | <table style="width: 100%; border-collapse: collapse; text-align: center; font-family: sans-serif; background: #fff; min-width: 800px;"> |
| | <thead> |
| | <tr style="background-color: #6366F1; color: white;"> |
| | <th style="padding: 14px; text-align: left; padding-left: 20px; border-radius: 8px 0 0 0;">Benchmark</th> |
| | <th style="padding: 14px; font-size: 1.1em;">Next 2.5 (4B) 🚀</th> |
| | <th style="padding: 14px;">Qwen 3.5 (4B)</th> |
| | <th style="padding: 14px;">Gemma-3 (4B)</th> |
| | <th style="padding: 14px;">Phi-4-Mini (3.8B)</th> |
| | <th style="padding: 14px; border-radius: 0 8px 0 0;">Llama-3.2 (3B)</th> |
| | </tr> |
| | </thead> |
| | <tbody style="color: #333;"> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: rgba(99, 102, 241, 0.08); font-weight: 600;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #4f46e5;">MMLU-Pro</td> |
| | <td style="padding: 12px; color: #10b981;">81.6%</td> |
| | <td style="padding: 12px;">79.1%</td> |
| | <td style="padding: 12px;">76.5%</td> |
| | <td style="padding: 12px;">78.2%</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;">MMLU-Redux</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">90.2%</td> |
| | <td style="padding: 12px;">88.8%</td> |
| | <td style="padding: 12px;">86.1%</td> |
| | <td style="padding: 12px;">87.5%</td> |
| | <td style="padding: 12px;">79.5%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: rgba(99, 102, 241, 0.08); font-weight: 600;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #4f46e5;">IFEval (Instruction)</td> |
| | <td style="padding: 12px; color: #10b981;">91.2%</td> |
| | <td style="padding: 12px;">89.8%</td> |
| | <td style="padding: 12px;">85.4%</td> |
| | <td style="padding: 12px;">88.1%</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;">HMMT (Reasoning)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">78.3%</td> |
| | <td style="padding: 12px;">74.0%</td> |
| | <td style="padding: 12px;">70.2%</td> |
| | <td style="padding: 12px;">72.8%</td> |
| | <td style="padding: 12px; color: #94a3b8;">--</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: rgba(99, 102, 241, 0.08); font-weight: 600;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px; color: #4f46e5;">LiveCodeBench v6</td> |
| | <td style="padding: 12px; color: #10b981;">58.4%</td> |
| | <td style="padding: 12px;">55.8%</td> |
| | <td style="padding: 12px;">51.0%</td> |
| | <td style="padding: 12px;">54.2%</td> |
| | <td style="padding: 12px;">45.1%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px;">TAU2-Bench (Agent)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">82.1%</td> |
| | <td style="padding: 12px;">79.9%</td> |
| | <td style="padding: 12px;">72.4%</td> |
| | <td style="padding: 12px;">75.0%</td> |
| | <td style="padding: 12px; color: #94a3b8;">--</td> |
| | </tr> |
| | </tbody> |
| | </table> |
| | </div> |
| | |
| | ### 👁️ Vision & Multimodal Edge |
| |
|
| | Next 2.5's visual cortex allows it to rival or beat proprietary nano-models from leading labs and larger 11B parameter open-weight models. |
| |
|
| | <div style="overflow-x: auto; box-shadow: 0 4px 6px rgba(0,0,0,0.05); border-radius: 8px; margin-top: 15px;"> |
| | <table style="width: 100%; border-collapse: collapse; text-align: center; font-family: sans-serif; background: #fff; min-width: 800px;"> |
| | <thead> |
| | <tr style="background-color: #4f46e5; color: white;"> |
| | <th style="padding: 14px; text-align: left; padding-left: 20px; border-radius: 8px 0 0 0;">Benchmark</th> |
| | <th style="padding: 14px; font-size: 1.1em;">Next 2.5 (4B) 🚀</th> |
| | <th style="padding: 14px;">Qwen 3.5 (4B)</th> |
| | <th style="padding: 14px;">Gemini-2.5 Flash-Lite</th> |
| | <th style="padding: 14px;">GPT-5-Nano</th> |
| | <th style="padding: 14px; border-radius: 0 8px 0 0;">Llama-3.2 (11B Vision)</th> |
| | </tr> |
| | </thead> |
| | <tbody style="color: #333;"> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px;">MMMU (General VQA)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">79.5%</td> |
| | <td style="padding: 12px;">77.6%</td> |
| | <td style="padding: 12px;">73.4%</td> |
| | <td style="padding: 12px;">75.8%</td> |
| | <td style="padding: 12px;">71.2%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: rgba(79, 70, 229, 0.05);"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px;">MathVision</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">76.8%</td> |
| | <td style="padding: 12px;">74.6%</td> |
| | <td style="padding: 12px;">52.1%</td> |
| | <td style="padding: 12px;">62.2%</td> |
| | <td style="padding: 12px;">50.5%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px;">OCRBench</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">86.5%</td> |
| | <td style="padding: 12px;">85.0%</td> |
| | <td style="padding: 12px;">82.5%</td> |
| | <td style="padding: 12px;">75.3%</td> |
| | <td style="padding: 12px;">74.1%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9; background-color: rgba(79, 70, 229, 0.05);"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px;">VideoMME (w/ sub)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">84.8%</td> |
| | <td style="padding: 12px;">83.5%</td> |
| | <td style="padding: 12px;">74.6%</td> |
| | <td style="padding: 12px;">71.7%</td> |
| | <td style="padding: 12px;">68.9%</td> |
| | </tr> |
| | <tr style="border-bottom: 1px solid #f1f5f9;"> |
| | <td style="padding: 12px; text-align: left; padding-left: 20px;">CountBench (Spatial)</td> |
| | <td style="padding: 12px; font-weight: bold; color: #10b981;">97.5%</td> |
| | <td style="padding: 12px;">96.3%</td> |
| | <td style="padding: 12px;">79.2%</td> |
| | <td style="padding: 12px;">80.0%</td> |
| | <td style="padding: 12px; color: #94a3b8;">--</td> |
| | </tr> |
| | </tbody> |
| | </table> |
| | </div> |
| | |
| | <p style="font-size: 0.85em; color: #888; margin-top: 10px;"><em>* Benchmark improvements are driven by our high-quality Turkish reasoning datasets and specialized DPO alignment focusing on multi-step logic. Empty cells (--) indicate scores not officially reported for that model.</em></p> |
| |
|
| | --- |
| |
|
| | ## 🚀 Quickstart & Usage |
| |
|
| | **Next 2.5** is fully compatible with the Hugging Face `transformers` ecosystem and modern serving frameworks like `vLLM` and `SGLang`. Because it is natively multimodal, you can pass images directly into the prompt. |
| |
|
| | ### Python (Transformers) |
| |
|
| | Make sure you have the latest `transformers`, `torch`, `torchvision`, and `pillow` installed. |
| |
|
| | ```python |
| | from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor |
| | from PIL import Image |
| | import torch |
| | |
| | model_id = "thelamapi/next2.5" |
| | |
| | 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 Next2.5, 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** | 4 Billion | |
| | | **Context Length** | 262,144 tokens natively (Extensible to 1M+ via YaRN) | |
| | | **Training Stage** | SFT + RLHF/DPO (Turkish + English focus) | |
| | | **Hardware** | Runs comfortably on consumer GPUs (e.g., RTX 3060/4060 with 8GB VRAM in FP16, or less via Quantization) | |
| | | **Capabilities** | Text Generation, Image Understanding, Video Summarization, OCR, Code Generation, Tool Use (Agentic) | |
| |
|
| | --- |
| |
|
| | ## 🎯 Ideal Use Cases |
| |
|
| | **Next 2.5 (4B)** strikes the perfect balance between high-end reasoning and hardware efficiency. It is perfectly suited for: |
| | * 🕵️ **Complex Document Analysis:** Upload massive PDFs or images of documents and extract structured, reasoned JSON outputs. |
| | * 🎓 **Educational Tutoring:** Its native `<think>` capabilities make it an excellent tutor that explains its mathematical steps to students. |
| | * 🤖 **Autonomous Agents:** Strong `Tool Calling` capabilities let you build desktop agents or web-browsing bots locally. |
| | * 🇹🇷 **Advanced Turkish NLP:** Finally, a mid-size multimodal model that understands Turkish idioms, grammar, and context as well as it does English. |
| |
|
| | --- |
| |
|
| | ## 📄 License & Open Source |
| |
|
| | Next 2.5 is released under the **Apache 2.0 License**. We support the open-source community and encourage developers to build commercial applications, conduct research, and innovate freely using this model. |
| |
|
| | --- |
| |
|
| | ## 📞 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 #e2e8f0; background: #f8fafc; border-radius: 8px;"> |
| | <p style="color: #475569; font-size: 15px; margin: 0;"> |
| | <strong>Next 2.5</strong> — Sınırları aşan görsel algı ve derin düşünme yeteneği. Türkiye'nin küresel yapay zeka vizyonu. 🌍 |
| | </p> |
| | </div> |
