thelamapi/next2.5

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

thelamapi/next2.5 is a 4.5 billion parameter Vision-Language Model (VLM) developed by Lamapi in Türkiye, built upon the Qwen 3.5-4B foundation. It features a native thinking mode using blocks for complex reasoning, a massive 262,144 token context length, and unified multimodal capabilities for images, documents, and video. This model is optimized for bilingual proficiency in Turkish and English, excelling in complex document analysis, educational tutoring, and autonomous agent tasks.

Loading preview...

Overview

Next 2.5 is a 4.5 billion parameter Vision-Language Model (VLM) developed by Lamapi in Türkiye, leveraging the Qwen 3.5-4B architecture. It has been extensively fine-tuned with culturally aware Turkish datasets and enhanced visual-spatial reasoning tasks, aiming to push the capabilities of mid-sized models.

Key Capabilities & Differentiators

  • Bilingual Proficiency: Tailored for flawless performance in both Turkish and English, with deep cultural and contextual understanding.
  • Native Thinking Mode: Utilizes <think> blocks to reason through complex logic, mathematics, and coding problems before generating responses.
  • Unified Multimodal Understanding: Natively processes images, performs OCR on documents, and analyzes hour-long videos.
  • Extended Context Length: Supports up to 262,144 tokens, making it suitable for large codebases or multi-document analysis.
  • Class-Leading Performance: Benchmarks show it outperforms other models in its parameter class (e.g., Gemma-3-4B, Phi-4-Mini) and rivals larger or closed-source edge models in vision and reasoning tasks, including MMLU-Pro (81.6%), IFEval (91.2%), and MathVision (76.8%).

Ideal Use Cases

Next 2.5 balances high-end reasoning with hardware efficiency, making it suitable for:

  • Complex Document Analysis: Extracting structured, reasoned JSON outputs from large PDFs or image-based documents.
  • Educational Tutoring: Leveraging its native <think> capabilities to explain mathematical steps.
  • Autonomous Agents: Building desktop agents or web-browsing bots due to strong tool-calling capabilities.
  • Advanced Turkish NLP: Providing a mid-size multimodal model with deep understanding of Turkish idioms, grammar, and context.

Finetunes
3 models