NVIDIA has rolled out a massive suite of open source AI models and local development tools. The update focuses heavily on running autonomous agents directly on personal computers and workstations without relying on cloud servers. But developers will need powerful graphics hardware to run the largest of these new releases.
Meta released Muse Glimmer, a 30 billion parameter model designed specifically for local coding and agent tasks. According to the NVIDIA local AI blog, this dense model runs on consumer hardware and achieves speeds over 200 tokens per second on an RTX 5090. The architecture uses hybrid attention to keep memory use low during long tasks. This lets developers process private documents, manage security keys, and run multi step workflows locally on 1 system without sending data to outside servers.
NVIDIA itself expanded its own offerings with Nemotron 3.5 Lightning, a 30 billion parameter mixture of experts model. This release is optimized for fast, specialized tasks and generates tokens up to 4 times faster than similar models in its class. Other developers also contributed to the rapid expansion of the open weights ecosystem this week. DeepSeek updated its DeepSeek V4 Flash model, while Alibaba launched Wan Animate 2, a 14 billion parameter model that animates static characters up to 26 times faster on an RTX 5090 than on competing silicon.
Video generation also got a major upgrade with the release of LTX 2.5. This open model improves visual quality, scene continuity, and prompt adherence using a custom text encoder. When running locally on an RTX 6000 graphics card, the model operates 20% faster while saving 40% of system memory. Creators can run these workflows directly on their own hardware through newly released ComfyUI setups.
To make running these models easier, Unsloth is launching Unsloth Desktop. This open source application allows users to both train and run AI models locally on 1 interface. For developers running massive models that exceed the capacity of 1 system, updates to the NVIDIA Sync app allow users to cluster multiple DGX Spark systems together. The tool automatically configures high speed connections between systems, reducing the cost of running advanced local agents to roughly 33% of previous setups by routing tasks dynamically to the best fit model.
