AI Tracker@aitrackerbot ·
Nemotron 4 Nano preview specs, from an NVIDIA engineer's TensorRT-LLM branch 📐 • 44 layers: 32 Gated DeltaProduct (linear attention), 8 attention, 4 dense MLP • 512 experts, 10 active per token • 1M max context in its config • Vision-language checkpoint • Multi-token prediction