Trendyol-LLM-Asure-12B is a 12-billion-parameter multimodal instruct model developed by Trendyol, built upon Gemma 3-12B. It is specifically optimized for structured instruction following with both text and image-text inputs, focusing on operational tasks in Turkish and English. Unlike general-purpose LLMs, its world knowledge is intentionally limited, instead being heavily tuned for e-commerce business tasks like summarization, QA, structured extraction, and controlled generation, while also offering efficient inference with reduced token consumption.
Loading preview...
Trendyol-LLM-Asure-12B Overview
Trendyol-LLM-Asure-12B is a 12-billion-parameter multimodal instruct model developed by Trendyol, based on Gemma 3-12B. This model is uniquely designed for operational task performance in Turkish and English, prioritizing structured instruction following over general encyclopedic knowledge. Its core differentiation lies in its specialized tuning for e-commerce business tasks, making it highly efficient for specific applications rather than broad conversational use.
Key Capabilities
- Multimodal (Vision + Text): Supports native image-text conversations, leveraging Gemma 3's multimodal features.
- Instruct-Optimized: Exclusively trained in instruct format, ensuring high prompt adherence and system-message compliance.
- Efficient Inference: Engineered for reduced token verbosity compared to its base model, leading to more efficient production inference.
- Bilingual: Demonstrates strong performance in both Turkish and English.
Good For
- E-commerce Business Tasks: Excels in summarization, paraphrasing, textual and visual Question-Answering (QA), structured extraction, controlled generation, text classification, and relevancy tasks specific to e-commerce.
- Structured Data Processing: Ideal for scenarios requiring precise instruction following and structured output from both text and image inputs.
- Production Environments: Optimized for lower token consumption and efficient inference, suitable for deployment in operational systems.
Limitations
The model's general world knowledge is intentionally limited, and it may generate inaccurate or misleading information outside its business-specific domain. Users should implement human oversight and application-specific safety testing due to potential biases and the risk of generating harmful content.