Orion by Meta

Microsoft’s new reasoning model, trained from scratch and optimized for code. It rivals the best models on SWE-Bench Pro and AIME 2025, while remaining medium-sized and computationally efficient

This 15-billion-parameter video model (developed by Alibaba) took first place in the Artificial Analysis rankings for video generation, ahead of Seedance 2.0. It generates text-to-video and image-to-video clips in 1080p with lip-sync in seven languages, all within a single unified pipeline

A 3D model that generates interactive 4D objects (cars with spinning wheels, automatic scripts) from textual instructions. Tokenization of 3D shapes, generation of shapes/scenes from text, and real-time character animation

An AI architecture with integrated long-term memory, selective updating based on a surprise signal, and context extended beyond 2 million tokens. MIRAS unifies transformers and linear RNNs thanks to optimized associative memory