Hugging Face

IBM has released its Granite 4.1 family of large language models, featuring 3B, 8B, and 30B dense decoder — only architectures.

Hugging Face has integrated DeepInfra as a new Inference Provider on its Hub, offering developers enhanced serverless AI inference options with a competitive pricing model and a diverse model catalog.

The financial and computational burden of evaluating advanced AI models, particularly agentic systems, has dramatically escalated, evolving into a primary compute bottleneck for the AI community, significantly impacting who can participate

NVIDIA has released Nemotron 3 Nano Omni, a highly efficient open-weights AI model that combines text, image, audio, and video processing to handle complex enterprise document analysis, audiovisual reasoning, and autonomous agent

NVIDIA and Siemens Healthineers have introduced an advanced AI model that directly processes raw ultrasound sensor data to dynamically estimate patient — specific speed of sound, utilizing novel edge computing hardware to improve real-time

OpenAI has released an efficient, 1.5 — billion-parameter open-source model on Hugging Face designed to securely detect and mask personally identifiable information across massive 128k — token document contexts, empowering developers

Google DeepMind, in collaboration with the Hugging Face platform, introduced the Gemma 4 family of open multimodal models.

A deep dive into the Sentence Transformers 5.4 update, which introduces native support for multimodal embedding and ranking models for processing text, images, audio, and video through a unified interface.

The TII research group introduced Falcon Perception and Falcon OCR - new compact models combining text and image processing in a single transformer with hybrid attention.

DeepSeek introduced its open models V4-Pro and V4-Flash with a hybrid attention architecture, radically reducing memory and computational resource consumption when working with a context of up to a million tokens, which opens new

IBM Research and Hugging Face released the VAKRA benchmark for evaluating AI agents in enterprise environments, published on April 15, 2026.

Hugging Face demonstrated the multimodal model Gemma 4 VLA, working entirely locally on NVIDIA Jetson Orin Nano Super. The model autonomously uses a webcam to obtain visual context, which

A detailed practical guide on integrating local AI models into browser extensions. It discusses the architectural features of the Manifest V3 standard, the strict separation of responsibilities between background processes and user

With the launch of Mythos and Project Glasswing, global organizations are beginning to realize a new stage in cybersecurity, where open tools and approaches play a critical role in software protection.

The Sentence Transformers library expands its functionalities, offering tools for training multimodal models that process text, images, audio, and video, which significantly improves their performance in specific tasks.

The launch of QIMMA represents a new stage in the evaluation of language models for the Arabic language, offering a systematic approach to validation and comparative analysis.

The new dataset from NVIDIA, allowing the creation of Korean AI agents, takes into account the cultural and demographic characteristics of Korean society.