Deep LearningMIT researchers have unlocked a revolutionary way to dynamically prune State-Space Models during training. By applying control theory to discard unobservable states, CompreSSM drastically reduces compute overhead without sacrificing performance.
Deep LearningNVIDIA has open-sourced AITune to automatically evaluate and select the fastest inference backend for PyTorch models. This new toolkit eliminates the optimization gap between research and production environments.
LLMDiscover how combining Nvidia's compact Nemotron-3-Nano model with WebGPU acceleration enables powerful, private, and real-time AI inference directly inside your web browser.
Deep LearningExplore the internal mechanics of MegaTrain, the trending deep learning framework that leverages extreme host memory offloading and asynchronous streaming to train massive foundation models on a single GPU.
LLMZhipu AI has open-sourced GLM-5.1, a massive 744-billion-parameter MoE model under the MIT license. Explore how it outperforms proprietary giants on software engineering benchmarks and what it means for the future of open-weight AI.
Deep LearningMIT researchers recently unveiled CompreSSM, a groundbreaking algorithm that compresses State-Space Models directly during training. By leveraging classical control theory, it delivers a massive 4x speedup and cuts compute costs without sacrificing model performance.
LLMLG AI Research has open-sourced EXAONE 4.5, a highly efficient 33B parameter multimodal model. It sets a new standard for open-source AI by outperforming proprietary systems in complex STEM evaluations and technical document parsing.
Machine LearningHugging Face has officially transferred the Safetensors project to the PyTorch Foundation. This strategic move establishes a vendor-neutral standard for model weights and finally eliminates the severe security risks associated with Python Pickle files.
Deep LearningHugging Face has officially transferred its Safetensors project to the PyTorch Foundation. This strategic move establishes a secure, high-performance standard for AI model weights and signals the beginning of the end for vulnerable Pickle files.
