Open modelsAn open-weight artificial intelligence model makes its trained parameter files available for others to download, run, inspect, or adapt under the model's license. dominate the roundup. GLM 5.3 Flash adds vision while combining sparse and linear attentionSparse attention reduces transformer computation by allowing each token to attend to a selected subset of other tokens instead of comparing with every token. for lower serving cost, the full GLM 5.3 weights approach frontier benchmarks, Qwen 3.8 Flash Next previews a more efficient long-context architectureA long-context artificial intelligence model can process substantially more tokens in one request, allowing it to work across long documents, conversations, or codebases., and Tencent's Hunyuan 4 Preview enters the same competitive range.
Video and 3D research also advanced. Block3D generates objects through parallel diffusion blocks, One Video One World reconstructs editable scenes from ordinary footage, Code World Model separates persistent world state from video rendering, and new MiniMax accelerators trade fewer denoising steps and sparse attention for much faster generation.
Google's Planetary Prediction Engine automates geographic data preparation and model building for questions such as disease and food-security risk. Gemini 3.5 Transcribe adds multilingual, low-latency speech recognition, while Orbit++ creates harder camera-motion benchmarks for 3D vision systems.
The robotics section shows humanoid machines running, jumping, playing games, and coordinating force with markedly better control than the previous year. Other projects demonstrate one-shot task imitation, local AI hardware with large unified memory, and more efficient reward-based fine-tuning for diffusion modelsA diffusion model generates or transforms data by learning to reverse a gradual noising process, often through a sequence of denoising steps..
Watch on YouTube
