NVIDIA released a draft Muse-Spark-Open-816B-A42B-DFlash2 model for speculative

On October 16, 2026, the model nvidia/Muse-Spark-Open-816B-A42B-DFlash2 appeared on HuggingFace with the image-text-to-text task. It is a DFlash 2 draft checkpoint for Muse-Spark-Open-816B-A42B, a multimodal mixture-of-experts model designed for long context and tool use. NVIDIA notes that the target model Muse-Spark-Open-816B-A42B was developed by a third party, while this draft was trained by NVIDIA to accelerate its inference via speculative decoding.

What it means

The model contains 4.4 billion parameters, and the largest weights file is model.safetensors at 8.2 GB. The weights are published in safetensors format, and launch is expected via Model Optimizer. There are no quantized builds on the platform yet. Under the license, the model is suitable for both commercial and non-commercial use: the NVIDIA Open Model Agreement applies, and the Meta Muse Model license is additionally mentioned. The developer does not specify VRAM requirements for running it. The model takes up a moderate amount of disk space, but it is a draft for accelerating the main 816B model rather than a standalone generative tool. You can follow the appearance of quantized builds and launch details in the CPU3D news feed.