NVIDIA engineers optimize Meta's Muse Spark Open model in TensorRT-LLM pull request

NVIDIA engineers have published a 10,236-line pull request in TensorRT-LLM detailing optimizations for Meta's Muse Spark Open model on NVIDIA GB300 Blackwell Ultra racks. The pull request reveals that the open-weight model features 816 billion total parameters, 256 experts, a 1-million token context window, and a 1.8-billion vision encoder. Additionally, several closed models from Meta recently went live on Google Cloud. Open article →