The model card lists an August 11, 2026 release date and says use is governed by the OpenMDW 1.1 model license.
The model has 30 billion total parameters and 3 billion active parameters per token. Its hybrid architecture combines Mamba-2, mixture-of-experts layers, and selected attention layers. NVIDIA positions the BF16 checkpoint for post-training, SFT, RL, distillation, domain adaptation, and creation of quantized or GGUF variants; it directs latency- and throughput-oriented deployment to a separate NVFP4 release.
The card lists single-GPU deployment on an 80GB H100 or A100. It validates 256K context on one H100, while 1M-token context is listed for GB200, B200, or an eight-H100 configuration. NVIDIA also documents configurable reasoning mode and a vLLM serving recipe.
NVIDIA reports benchmark results under its NeMo Gym and NeMo Evaluator harnesses and notes that results may differ from vendors' self-reported figures. The source does not establish independent benchmark performance, real-world cost, throughput, or performance on a particular workload.