The post describes Muse Spark Open as a mixture-of-experts coding model with 816 billion total parameters, 42 billion active parameters per token, and a one-million-token context window.
NVIDIA says NVFP4 checkpoints are optimized for Blackwell Ultra, and points developers to SGLang, vLLM, TensorRT LLM, and NVIDIA Dynamo deployment options. It also says the weights can be downloaded from Hugging Face and that a downloadable NVIDIA NIM is available for production setup. NVIDIA describes GB300 NVL72 as a 72-GPU NVLink domain.
For agentic workloads, NVIDIA claims up to 50,000 tokens per second per GPU and up to 600 tokens per second per user on four Blackwell GPUs with speculative decoding. The source does not establish an independently reproduced benchmark, the exact workload and concurrency behind every result, comparative performance on other hardware, or deployment pricing.