Nemotron-H-47B

Nemotron-H-47B-Base is NVIDIA's efficiency play in the open-model space: a hybrid architecture that replaces most self-attention layers with Mamba-2 layers, keeping only a handful of attention layers for the things recurrence handles poorly. The 47B variant was produced from the 56B flagship by MiniPuzzle, NVIDIA's own prune-and-distill compression recipe, and NVIDIA reports it matches the 56B on accuracy while inferring about 20% faster. The family paper (arXiv:2504.03624) positions these models against Qwen2.5-72B and Llama-3.1-70B at up to 3x the inference speed.

The 56B was the first Nemotron model fully pretrained in FP8, on 20 trillion tokens, and the 47B inherits that training. There is also a 128K reasoning variant for agent-style workloads. The important caveat is licensing: these weights ship under NVIDIA's Internal Scientific Research and Development license, which is research-only, not a commercial open license. For teams that can live with that, it is one of the most serious hybrid SSM models available; for products, the license is the blocker.

• NVIDIA-reported: Nemotron-H-47B-Base matches the 56B model on accuracy while inferring ~20% faster via MiniPuzzle pruning and distillation (provider claim, arXiv:2504.03624)
• NVIDIA-reported: family competitive with Qwen2.5-72B-Base and Llama-3.1-70B-Base at up to 3x faster inference (provider claim)
• No independent benchmark table verified