NVIDIA Nemotron 3.5 Lightning 30B-A3B

Nemotron 3.5 Lightning is NVIDIA's 30B mixture-of-experts language model with roughly 3B active parameters, released under the OpenMDW-1.1 license. The architecture is a Mamba-2 / MoE / attention hybrid: 52 layers alternating Mamba-2 state-space blocks, MoE blocks (128 routed experts, 6 active per token, plus 1 shared expert) and periodic full-attention layers, with an SSM state size of 128 and a float32 state cache. Context length reaches up to 1M tokens — NVIDIA deploys it at 256K on a single H100. Reasoning is switchable on and off through the chat template, and the model covers English plus Spanish, French, German, Italian and Japanese as well as code. The BF16 repository linked here holds the full-precision reference weights; NVIDIA also publishes NVFP4 variants, and DSpark/DFlash speculative-decoding checkpoints. Runnable on mainline llama.cpp (arch 'nemotron_h_moe', GGUF published by ggml-org), through the Ollama library as nemotron-3.5-lightning, and on Apple silicon via mlx-lm.