ogulcanaydogan/Turkish-LLM-32B-Instruct

TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Mar 21, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Turkish-LLM-32B-Instruct by ogulcanaydogan is the largest openly available 32.8 billion parameter Turkish-enhanced language model, fine-tuned from Qwen2.5-32B-Instruct. It excels in Turkish STEM, reasoning, and general knowledge tasks, demonstrating significant improvements on MMLU-TR (+2.71) and XCOPA-TR (+1.00) benchmarks. This model is specifically optimized for high-quality Turkish instruction following, making it ideal for applications requiring robust Turkish language understanding and generation.

Loading preview...

Turkish-LLM-32B-Instruct: The Largest Open-Source Turkish LLM

Turkish-LLM-32B-Instruct, developed by ogulcanaydogan, is the largest openly available 32.8 billion parameter language model specifically fine-tuned for the Turkish language. Built upon the Qwen2.5-32B-Instruct base model, it leverages QLoRA with a meticulously curated 173K Turkish instruction dataset, emphasizing quality over quantity in its iterative development.

Key Capabilities

  • Superior Turkish Performance: Achieves 67.89% on MMLU-TR, a significant +2.71 point improvement over its base model, and a +1.00 point improvement on XCOPA-TR for causal reasoning.
  • Optimized for STEM & Reasoning: Shows strong gains in categories like College Computer Science (+7.1), Logical Fallacies (+5.6), and College Mathematics (+5.0).
  • Efficient Local Inference: GGUF quantizations (Q4/Q5/Q8) are available, enabling deployment on consumer hardware.
  • Iterative Dataset Engineering: Benefits from a refined dataset (v7.1) that prioritizes data quality and evaluation alignment, leading to enhanced performance.

Good for

  • Applications requiring advanced Turkish language understanding and generation.
  • Tasks involving Turkish STEM, logical reasoning, and general knowledge.
  • Developers seeking the largest and most capable open-source Turkish instruction-tuned model.
  • Local deployment scenarios leveraging GGUF quantizations for efficient inference.