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
🇹🇷 Tailored in Türkiye: Flawless bilingual proficiency (TR/EN) with deep cultural and contextual awareness.
🧠 Native Thinking Mode: By default, it uses <think>...</think> blocks to reason through complex logic, math, and coding tasks before answering.
👁️ Unified Vision-Language: Natively understands images, documents (OCR), and hour-long videos without breaking a sweat.
📈 Class-Leading Performance: Outperforms heavyweights in its parameter class (Gemma-3-4B, Phi-4-Mini) and even rivals closed-source edge models like GPT-5-Nano.
📚 Massive Context Limit: Supports up to 262,144 tokens natively, perfect for massive codebases or multi-document analysis.
📊 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)
Benchmark
Next 2.5 (4B) 🚀
Qwen 3.5 (4B)
Gemma-3 (4B)
Phi-4-Mini (3.8B)
Llama-3.2 (3B)
MMLU-Pro
81.6%
79.1%
76.5%
78.2%
68.4%
MMLU-Redux
90.2%
88.8%
86.1%
87.5%
79.5%
IFEval (Instruction)
91.2%
89.8%
85.4%
88.1%
77.4%
HMMT (Reasoning)
78.3%
74.0%
70.2%
72.8%
--
LiveCodeBench v6
58.4%
55.8%
51.0%
54.2%
45.1%
TAU2-Bench (Agent)
82.1%
79.9%
72.4%
75.0%
--
👁️ 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.
Benchmark
Next 2.5 (4B) 🚀
Qwen 3.5 (4B)
Gemini-2.5 Flash-Lite
GPT-5-Nano
Llama-3.2 (11B Vision)
MMMU (General VQA)
79.5%
77.6%
73.4%
75.8%
71.2%
MathVision
76.8%
74.6%
52.1%
62.2%
50.5%
OCRBench
86.5%
85.0%
82.5%
75.3%
74.1%
VideoMME (w/ sub)
84.8%
83.5%
74.6%
71.7%
68.9%
CountBench (Spatial)
97.5%
96.3%
79.2%
80.0%
--
* 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.
🚀 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
1from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor
2from PIL import Image
3import torch
45model_id ="thelamapi/next2.5"67model = AutoModelForCausalLM.from_pretrained(model_id)8processor = AutoProcessor.from_pretrained(model_id)# For vision.9tokenizer = AutoTokenizer.from_pretrained(model_id)101112# Create a message in chat format13messages =[14{"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."}]},1516{17"role":"user","content":[18{"type":"text","text":"Write a highly optimized Rust function to calculate the Fibonacci sequence using memoization"}19]20}21]2223# Prepare input with Tokenizer24prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)25inputs = processor(text=prompt, return_tensors="pt")2627# Remove 'mm_token_type_ids' if it's not needed for text-only generation28if"mm_token_type_ids"in inputs:29del inputs["mm_token_type_ids"]303132# Output from the model33output = model.generate(**inputs, do_sample=True, temperature=0.7, max_new_tokens=128)34print(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.
