© ROOT-NATION.com - Use of content is permitted with a backlink.
As part of its Local AI Month, NVIDIA, together with its partners, unveiled new open-source models and tools for building and running autonomous agents directly on users’ devices.
See also: AERONAUT – everything that flies above the ground: aviation, UAVs and drones, rockets, and space
Meta has released the Muse Glimmer model, featuring 30 billion parameters and a context window of over 120,000 tokens, designed for writing code and autonomously performing complex tasks. Thanks to optimizations from NVIDIA, on a PC with an RTX 5090 graphics card, it generates over 200 tokens per second, enabling local processing of private files and documents without connecting to the internet.
The multimedia sector has received the LTX-2.5 model for creating high-quality videos with improved frame sequences and character animations. Meanwhile, Alibaba introduced Wan-Animate-2, a 14-billion-parameter model that transfers facial expressions from video to static images and runs 26 times faster on the RTX 5090 compared to the Apple M3 Ultra. Developers also now have access to the MiniMax-H3 video generator with 33 billion parameters and the Cosmos 3 Edge robotics model.
Other solutions include the Inkling-Small multimodal model from Thinking Machines Lab with 276 billion parameters, the updated DeepSeek-V4-Flash with 1 million token context, and the Laguna S 2.1 coding agent from Poolside AI.
The ecosystem has also expanded with new software. Unsloth introduced the open-source Unsloth Desktop application, which allows users to both train and run models locally. To scale computations, NVIDIA updated the NVIDIA Sync tool, adding Cluster Assistant to combine DGX Spark systems via ConnectX-7 ports into a single cluster. At the end of August, the NVIDIA Sync Resource Monitor and support for the Google Chrome browser on Linux ARM64 will be added.