NVIDIA's new 30B hybrid Mamba-Transformer MoE activates just 3B parameters, runs on a single H100, and handles up to 1M tokens of context.
document fetched 2026-10-26 02:56