AMD Unveils Instella-VL-1B Vision Language Model

AMD has announced its first vision language model, Instella-VL-1B, which is trained on AMD's Instinct MI300X GPUs. This model is part of the Instella family of language models introduced by AMD in March 2025. Instella-VL-1B is a multi-modal model featuring 1.5 billion parameters, combining a vision encoder with 300 million parameters and a language model with 1.2 billion parameters.

The model was developed using datasets such as LLaVA, Cambrian, and Pixmo, and was further enhanced with document-related datasets like M-Paper and DocStruct4M. With a new pre-training dataset of 7 million examples and a supervised fine-tuning dataset of 6 million examples, Instella-VL-1B outperforms similarly sized open-source models on general visual language tasks and OCR-related benchmarks.

AMD has made Instella-VL-1B fully open-source, sharing not only the model weights but also detailed training configurations, datasets, and code. This initiative underscores AMD's commitment to advancing open-source AI technology in the field of multimodal AI.