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
🇹🇷 Perfected in Türkiye: Fine-tuned with cultural nuance, ensuring natural, fluent, and highly accurate Turkish responses.
💨 "Air" Speed & Efficiency: Only 2 Billion parameters. Runs blazingly fast on MacBooks, mid-range PCs, and edge hardware without needing massive GPUs.
🧠 Native Thinking Mode: Despite its small size, it leverages Chain-of-Thought (<think>) to logically deduce answers before speaking.
👁️ Full Vision-Language Support: Analyzes images, reads documents (OCR), and understands visual context just like heavier models.
📚 Massive Context: Supports a staggering 262,144 tokens natively—perfect for summarizing long PDFs or reading extensive codebases locally.
📊 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
Benchmark
Next2-Air (2B)
Qwen 3.5 (2B)
Gemma-2 (2B)
Llama-3.2 (3B)
MMLU-Pro (Thinking)
68.2%
66.5%
54.1%
68.4%
MMLU-Redux
82.1%
79.6%
75.3%
79.5%
IFEval (Instruction)
82.5%
78.6%
75.8%
77.4%
TAU2-Bench (Agent)
52.4%
48.8%
--
--
👁️ Multimodal & Vision Edge
Next2-Air features a highly capable visual encoder, allowing it to process spatial intelligence, OCR, and document understanding tasks efficiently.
Benchmark
Next2-Air (2B)
Base Qwen3.5-2B
MMMU (General VQA)
66.5%
64.2%
MathVision
78.1%
76.7%
OCRBench
86.0%
84.5%
VideoMME (w/ sub)
77.8%
75.6%
* 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.
🚀 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
1from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor
2from PIL import Image
3import torch
45model_id ="thelamapi/next2-air"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 Air, 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
2 Billion (Ultra-Lightweight)
Context Length
262,144 tokens natively
Hardware
Optimized for Edge devices, MacBooks (MLX), Consumer GPUs, and low-VRAM environments.
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.
