Artificial Intelligence Community Platform Hugging Face‘in Smolvlm-256m and smolvlm-500m that can analyze the team, images, short videos and texts Published models. The team claims that these models are the smallest artificial intelligence models.
According to the team, the models are designed to work well on limited devices such as laptops with less than about 1 GB of RAM. The smallest version uses 16 examples in seconds, while a 64 -party party uses only 15 GB of RAM. In this sense, we can say that developers who try to process a large amount of data very cheaply can choose SMOLVLM-256M and Smolvlm-500M models.
Both models can fulfill tasks such as answering the questions about PDFs and items in addition to identifying images or video clips. It should be noted that text and graphs scanned within the scope of PDFs can also be processed.
The Huging Face team used The Cauldron and Docmatix to train models. The Cauldron consists of 50 high quality images and text data clusters. Docmatix is a file scanning set matched with detailed subtitles. In the meantime, it is worth noting that both of them were created by the M4 team that develops multimodal artificial intelligence technologies of Huging Face.
The team says that both SMOLVLM-256M and SMOLVLM-500M performs better than the Idefics 80B model, a system of 300 times larger in various comparisons such as AI2D. Let’s add that the AI2D criterion tested the ability of models to analyze the primary school level of science diagrams. SMOLVLM-256M and SMOLVLM-500M can be downloaded on the web and from Huging Face under Apache 2.0 license. At this point, the models can be used without restriction.
Source link: https://webrazzi.com/2025/01/24/hugging-face-in-yeni-kucuk-yapay-zeka-modelleri-smolvlm-256m-ve-smolvlm-500m/