smolagents now supports vision language models (VLMs) by allowing images to be passed to the agent at start or dynamically via callbacks, enabling use cases like autonomous web browsing.
Hugging Face announces two new small vision-language models: SmolVLM-256M (256M parameters) and SmolVLM-500M (500M parameters), which are the smallest VLMs in the world, with improved efficiency and…
Hugging Face and FriendliAI announced a partnership integrating FriendliAI Endpoints into the Hugging Face Hub as a deployment option, enabling one-click deployment of generative AI models.
OpenAI published a system card for Operator that describes safety mitigations against prompt injection and jailbreaks, privacy protections, external red teaming, and ongoing safety work.
OpenAI expresses enthusiasm for its vision of AGI infrastructure and seeks to partner with firms across the data center infrastructure landscape, including power, land, construction, and equipment.
Alibaba's Qwen team released two open-source models, Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, which support a context length of up to 1 million tokens.
Alibaba releases Qwen2.5-VL, a new vision-language model, and opens base and instruct models in three sizes (3B, 7B, 72B) on Hugging Face and ModelScope.
The post discusses the Mixture-of-Experts (MoE) architecture, which is a technique for scaling model parameters, and describes how an MoE layer consists of a router and a group of experts.
LG AI Research presented their paper 'AGATa: Attention-Guided Augmentation for Tabular Data in Contrastive Learning' at the NeurIPS 2024 TRL Workshop, introducing a novel attention-guided data…